Developer tools
Latest stories covering Developer tools, with summaries and links to related coverage.
45 stories · 45 reports · page 1 of 1
The topic overview is temporarily unavailable. You can still browse coverage below.
Connected topics are temporarily unavailable. You can still browse stories below.
Latest stories
There are now 5 inference providers for abliterated DeepSeek
tool
1 min read
AllModels routes abliterated language models across providers, offering automatic fallbacks when a selected provider is unavailable. The service uses modality-aware selection to match routing with model capabilities. AllModels also provides a single place to compare model availability and token prices across providers.
Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)
news
19 min read
Cloudflare remediated a tenant-isolation flaw in its Workers AI deployment of Z.ai’s GLM-5.3 model after security researcher Gadi Evron of Knostic reported it through HackerOne on September 8, 2026. From the model’s August 27 launch until September 9, 155 responses to prompts exceeding 681,280 tokens contained data from other customers’ recent requests to the same model. Cloudflare found no evidence of deliberate exploitation and said the affected request timing and volume matched ordinary long-context use. The flaw combined a Cloudflare configuration error with a bug in the SGLang inference engine’s disaggregated prefill/decode processing. GLM-5.3 accepted a 1,048,576-token context window, but its prefill KV-cache pool held only 681,280 tokens. For over-limit prompts, SGLang replaced the request with a one-token placeholder yet passed it to decode as completed; decode then read stale KV-cache pages left by earlier requests and could generate responses containing their data. Cloudflare made GLM-5.3 private on September 9, corrected the pool to about 2.04 million tokens, deployed an engine patch, and restored public access that day. It applied related hardening to all SGLang deployments. SGLang released fixes in versions 0.5.20 through 0.5.22; Cloudflare recommends operators of disaggregated SGLang deployments use v0.5.22 or later.
The Git Forge for Agent-Human Collaboration
tool
10 min read
Entire has launched a Git-compatible code-hosting platform designed for teams working with coding agents. Its Trail workflow replaces conventional pull-request timelines with agent-generated evaluation scores, findings and merge gates displayed for every code push. Trails can span multiple repositories, run configurable agents for code quality, security and “slop,” and require specified agent findings, human approvals and CI checks before merges. Entire says its review agents use Claude Code and Codex models in isolated Daytona and E2B cloud sandboxes. The platform natively hosts repositories or mirrors GitHub repositories, supports replication across regions for data residency, and is available in the US, EU, Australia and India. Entire says its distributed object store and transactional ref database enable pushes up to 89 times faster than GitHub and six times faster than Cursor Origin, based on its open-source benchmarks. Buildkite and Depot integrations provide CI checks; failed builds can block merging. Entire also introduced Continuous Memory through open-source, local-first CLI plugins. entire-graph locally analyzes repositories with tree-sitter to map code definitions, callers, types, routes and change impact; Entire says testing showed token savings of up to 71%. entire-brain indexes that code map alongside agent sessions recorded in Git history, allowing agents and reviewers to retrieve prior technical context and decisions.
Anthropic S-1
news
1 min read
An SEC filing link identifies Anthropic, PBC and an S-1 document labeled “Q2.” The source provides no filing details, financial figures, dates, business information, or statements from Anthropic. The link was posted by SiNTEx and had a score of 1,064 points and 989 comments at the time reflected in the source.
Fixing Tool-Call Accuracy in vLLM
research
8 min read
Ubicloud enabled constrained decoding in vLLM to enforce valid tool-call schemas for Kimi and DeepSeek models, reporting an increase in K2VV ToolCall benchmark schema accuracy from 92% to 100%. The company had observed that Kimi K2.6 served through vLLM 0.28 produced the expected tool-call schema 92% of the time, meaning roughly two of every 25 calls were malformed. Invalid calls can add error messages to an agent’s context, reduce output quality, and require retries that increase costs. Tool calls require an LLM to emit structured arguments matching a declared JSON schema, so an agent harness can invoke functions such as web-page fetching. Constrained decoding prevents the inference engine from selecting tokens that would violate the required tool-call structure or JSON schema during generation. vLLM uses structural tags from XGrammar to restrict valid tokens at each stage, including tool names, call indexes, and arguments. Although vLLM added constrained decoding in version 0.25, it does not enable strict tool decoding by default for all requests because support is model-specific and can cause problems for some models. vLLM later added the server-side --tool-strict-level flag in version 0.31. Moonshot AI recommends constrained decoding for Kimi, and Ubicloud deployed it after validating 100% schema accuracy across its served model versions.
Writing Is Closer to Thought Than Code
opinion
11 min read
Daniel Miessler argues that written articulation of a project’s intended outcome should be the primary reference for AI-assisted development, rather than source code. He disagrees with developer Matt Pocock’s view that code is the central project artifact, contending that code records what was built but cannot fully capture why it was built or what people originally wanted. Miessler says many AI failures in 2026 stem from incomplete communication of intent rather than poor execution by AI systems. Miessler says he adopted an “Ideal State Artifact” (ISA) at the end of 2025 to document projects. An ISA includes a prose description of the desired state, constraints, dated decisions made during development, and Ideal State Criteria (ISCs): atomic, testable requirements paired with evaluations. Evaluations can range from basic assertions to multi-stage processes involving code execution, question sets, rubrics, tournaments, or an LLM judge. The ISA is intended to preserve requirements separately from a particular implementation, allowing teams to discard and rebuild inadequate code against the same criteria. Miessler describes two iteration loops: refining the written description as intent becomes clearer, and testing implementations against that description. He argues that increasingly capable AI shifts the main risk from implementation errors toward gaps between a person’s actual intent and the intent captured in writing.
PaperPilot – Zero-Cloud, Pure-Rust PDF Engine with WASM and MCP
tool
3 min read
PaperPilot is a Rust-based local PDF and document toolkit available as an MCP server, command-line interface, browser-based WebAssembly tool, Tauri desktop application, and local REST API. It supports PDF merging, splitting, rotation, page deletion and reordering, page extraction, booklet imposition, N-up grids, compression, text extraction, OCR, watermarking, signature validation or removal, PII scanning, deep sanitization, and Excel-to-JSON or CSV conversion. The MCP server connects through local standard input/output to Codex, Claude Code, Kimi Code, Pi, Claude Desktop, Cursor, and other MCP clients, allowing agents to execute PDF operations. The local REST gateway runs on localhost port 7823 and can be called from Python, Node, or n8n. Browser processing keeps files in browser memory rather than uploading them, while the project states that its tools make no cloud calls and require neither Chromium nor a C++ runtime. PaperPilot requires Rust 1.75 or later; desktop development also requires Node.js and pnpm. Its Apache 2.0 Community Edition includes the CLI, MCP server, REST gateway, WASM engine, and desktop app. Planned releases include PDF conversion to Markdown, HTML, JSON, Word, Excel, and PowerPoint, followed by encryption, signing, redaction, forms, rendering, repair, comparison, PDF/A support, and Bates numbering. A separate Enterprise Edition is planned with SSO, role-based access control, audit logs, and DLP.
Microsoft Decision-1: 9B Qwen Model Launch Explained
tool
21 min read
Microsoft introduced Microsoft-Decision-1 on October 9, 2026, a specialized model that scores predefined answer choices rather than generating text. Built by post-training Alibaba’s 9-billion-parameter open-weight Qwen3.5-9B model, it returns calibrated probability scores for tasks including routing, classification, prioritization, verification and workflow control. It supports text inputs of up to 32,000 tokens but does not provide written explanations or process images, audio or video. Microsoft said future versions will be rebased on Microsoft AI and OpenAI models. The model launched in public preview through Microsoft Foundry and became available on OpenRouter as microsoft/microsoft-decision-1. Its weights are not distributed, limiting use to hosted APIs. OpenRouter and secondary reports list pricing at $0.042 per million input tokens with no output-token charge. Microsoft said Decision-1 achieved 83.5% accuracy across 36 benchmarks containing roughly 150,000 held-out questions, and reported approximately 35 times lower P50 latency than GPT-6 Sol and 4.5 times faster performance than Quyet-1.0-Large; those results have not been independently reproduced. Microsoft also reported 98.7% perturbation stability and said it is testing the model internally for incident response, quality control and scientific discovery. Xbox Research reportedly used it to categorize more than 10,000 feedback items into predefined themes.
SIMD-accelerated computer vision on a $2 microcontroller
research
5 min read
A SIMD-accelerated implementation of the FAST feature detector runs on Espressif’s ESP32-S3 microcontroller, using the chip’s 128-bit vector instructions to process 16 eight-bit pixels per vector register. The implementation processes a 320×240 QVGA frame in about 6 ms and increased measured throughput from 5.1 megapixels per second to 11.2 megapixels per second, a roughly 220% improvement over the reference implementation. The stated performance is sufficient to process a 30-frame-per-second VGA stream. The ESP32-S3 has a dual-core CPU clocked up to 240 MHz, Wi-Fi and Bluetooth Low Energy radios, and a limited SIMD instruction set. FAST identifies image features by comparing pixels around a center pixel; its pre-test samples pixels in the cardinal directions and looks for at least three whose intensity differs from the center by more than a threshold. The implementation handles the ISA’s lack of direct misaligned reads by extracting slices from adjacent vector registers. It also compensates for signed-only eight-bit SIMD comparisons and saturating arithmetic by XORing each unsigned pixel value with 0x80, mapping the 0–255 range into the signed −128–127 range without subtraction.
The State of Context Engineering
opinion
18 min read
Context engineering has evolved from manually selecting prompt information to building software harnesses that retrieve, filter, prune, summarize, and isolate information for AI models. Dave Breunig argues that the central task is deciding which behavior should be enforced by deterministic code and which should be delegated to probabilistic model calls. Long context windows, which expanded from roughly 128,000 tokens to millions beginning with Gemini 1.5, can accommodate more data but also more errors and conflicting signals. Breunig identifies context poisoning, distraction, confusion, and clash as major failure modes. Google reported that Gemini 2.5 playing the original Pokémon Red repeatedly pursued tea, an objective from Pokémon FireRed embedded in its training data, despite the older game requiring lemonade. Research by Carolina Camassa and Derek Shiller found that contradictory examples in a conversation can increasingly override explicit instructions. Google also reported that Gemini 2.5 began repeating historical actions rather than making new plans once Pokémon-agent context grew well beyond 100,000 tokens. Harness techniques include retrieval-augmented generation, selective tool loading, fresh-context subagents, history pruning, external storage, and compaction. Breunig says agentic models introduce tool-environment biases and “context inertia,” in which agents persist with an unproductive hypothesis. Anthropic’s Memento tests and research from Microsoft and UC Berkeley found that periodically clearing conversational context while retaining work artifacts can restore exploration and improve results.
What Adobe v ArtCraft teaches us
opinion
7 min read
The writer argues that low-cost AI coding tools from companies including OpenAI and Anthropic could erode the software businesses expected to become their largest customers. The argument is that established software companies have historically charged recurring fees partly because complex products required large engineering teams and years of investment to build. If AI lets hobbyists or small teams create comparable alternatives cheaply and release them for free or at low prices, subscription revenue at incumbent software vendors could decline rather than shift to AI labs. The writer characterizes $20-per-month coding plans and $200 premium plans as subsidized pricing, claiming heavy users may consume more computing capacity than they pay for. Amazon, Alphabet and Meta are cited as planning roughly $500 billion in 2026 capital spending, much of it for AI data centers. The writer says AI labs face a pricing dilemma: higher enterprise-style prices could push users toward free open-weight models, while low prices may fail to cover infrastructure costs. Meta's February 24, 2023 release of Llama and subsequent open-weight releases from Chinese labs are presented as making durable scarcity difficult. The writer contends that OpenAI and Anthropic lack Nvidia's hardware business, Meta's larger business lines, and vertical integration, leaving their business models exposed if AI makes software development abundant and inexpensive.
We should be more ambitious with the software we build
opinion
26 min read
Dan Luu argues that increasingly capable coding models let individuals build software projects that were previously too costly in time or expertise, while warning that fast generation does not ensure correctness, performance, or maintainability. He says he built a customizable regex engine, a faster ripgrep fork for his workloads, and a Rust-related build-speed tool in a low single-digit number of hours with AI assistance; he estimates these projects would previously have taken a year or more to reach “works for me” quality. He says newer models have also made it easier to implement techniques from papers and codebases in specialized areas such as board-game AI. Luu reports that GPT-6 Astra passed registered zstd implementation tests more than 99% of the time in his evaluations and cost about $1 per implementation, though he says more ambitious tasks such as IMAP and Pandoc remained useful tests. He advocates building systems with self-checking properties, including fuzzing, property-based testing, consistency checks, and formal verification, rather than relying on one-shot AI output. Yossi Kreining cautions that most users should moderate their ambitions because many problems lack clear pass/fail criteria and AI-generated systems can become opaque, brittle, and hard to maintain. Luu also says pgrust is not yet suitable as a primary production database, citing founder Michael Malis’s September 17, 2026 statement that its query engine still has regressions and unsupported complex queries.
Dream-1
tool
10 min read
Dream Labs has launched Dream-1, a World Action Model available through a public API that it says can control previously unseen bimanual robots with parallel grippers without robot-specific fine-tuning or demonstrations. The company emerged from stealth with $90 million in seed funding and positions the system as a way to make newly deployed, lower-cost robots useful before they have accumulated their own training data. Dream-1 predicts end-effector poses—gripper positions, orientations, and open or closed states—rather than only robot-specific joint angles. A robot’s inverse-kinematics solver converts those poses into joint commands. The model uses a head camera and wrist cameras, and Dream Labs says its video-and-action training enables it to generalize across robots whose bodies and grippers differ. The company reported tests on multiple bimanual parallel-gripper systems, including the 18-year-old PR2 robot. Robocurve independently evaluated Dream-1 before launch, according to Dream Labs. Dream Labs says the model costs $3 per hour and Dream-1-flash costs $2 per hour; each API call returns 1.6 seconds of actions. For robots where zero-shot performance is insufficient, its Fine-tuning-as-a-Service adapts the model using robot-specific data. Dream Labs reports five minutes of data improved results for two partner robots, while other configurations required 77 to 98 minutes.
Can we replace a $150k Looker with a bucket of HTML files?
opinion
13 min read
A three-engineer data platform team at a roughly 500-person company built an internal AI-generated dashboard service that hosts HTML files, runs live Snowflake queries, and has reached at least as many weekly users as the company’s Looker deployment. The company pays about $150,000 annually for roughly 300 Looker seats, while the internal service grew from 23 weekly viewers in March to 136 in June. Users generate free-form dashboard pages with Claude, publish them through an internal connector, and share them behind single sign-on. The service initially exposed serious security risks because unreviewed JavaScript ran on the same origin as user sessions and every employee could access every dashboard. The team later added a remote MCP connector, draft-and-publish workflow, sandboxed iframes on separate origins, strict content-security policies, one-hour viewer-specific content tokens, and metadata scanning that can flag or block risky scripts and WebRTC connections. Snowflake OAuth supplies each viewer’s identity and data permissions, so sharing a dashboard does not grant access to underlying data. The team reports 325 commits, about 74,000 lines of code including 20,000 test lines, and roughly $6,400 in Claude token costs over eight months. It estimates annual operating costs at about $20,000, including AWS and ongoing Claude use. Its analyst agent uses about 1 million tokens per dashboard, with reported costs falling from about $2.50 to $0.50 after moving to Opus 5.5.
Kimi K3.1 Pro and K3.1 Pro Highspeed
tool
15 min read
Kimi has introduced Kimi K3.1 Pro, a post-trained successor to Kimi K3 and its most capable 3-trillion-parameter-class open-weight model, with a one-million-token context window. Kimi K3.1 Pro Highspeed is also available, offering two to three times the standard model’s speed. Both models are available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API; Kimi says it will release the full weights by November 4, 2026. Kimi reports that K3.1 Pro scored 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5, and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, up from K3’s 55.95%. It reported a 92.7% tool-enabled score on BenchCAD and an OfficeQA Pro v2 score of 64.4. Kimi also said its internal stress test measured a claim-level hallucination rate of 8.3%, down from 23.4% for K3; those figures are not daily-use error rates. For the open-weight release, Kimi removed cyber-related data from mid-training and filtered cyber tasks from reinforcement-learning environments. On its xdev-bench-hard vulnerability benchmark, Kimi said a checkpoint without those measures achieved 90.5% pass@3, while the final model achieved 81.0%. Kimi said the controls reduced cyber capability while retaining general coding usefulness.
Swedish government offices to ban ChatGPT and Claude
news
2 min read
Regeringskansliet blockerar från och med måndag AI-tjänsterna ChatGPT och Claude på anställdas tjänstedatorer. Syftet är att styra medarbetare till AI-tjänster som organisationen har godkänt, däribland Microsoft Copilot. Digitaliseringsdirektör Mikael Wettercrantz säger på Regeringskansliets intranät att användning av AI-tjänster utan avtal kan få konsekvenser för informationssäkerheten. Regeringskansliet har samtidigt, efter en översyn, skärpt säkerheten i Copilot. Presstjänsten har inte lämnat detaljer om ändringarna med hänvisning till säkerhetsskäl. Vissa anställda kommer fortsatt att kunna använda en särskild plattform för AI-assistenter som Regeringskansliet har utvecklat, där ChatGPT och Claude finns tillgängliga. Användningen av DeepSeek hade redan begränsats. Flera svenska myndigheter har vidtagit eller överväger liknande åtgärder. Skatteverket och Statens institutionsstyrelse har infört spärrar, medan Kronofogden bereder frågan inför ett beslut. Trafikverkets internrevision rapporterade tidigare under året om utbredd och riskfylld användning av omkring tio AI-tjänster, inklusive hantering av känslig information och AI-genererade svar.
OpenShell – A safe, private runtime for fleets of autonomous AI agents
tool
4 min read
NVIDIA’s OpenShell is a runtime for fleets of autonomous AI agents that lets administrators define and enforce policies governing files, system calls, network access, APIs, packages, and credentials. Each agent runs in an isolated sandbox, while kernel-level controls enforce restrictions on file access and system calls and check network connections before they leave the sandbox. OpenShell supplies credentials only on requests sent to approved endpoints, so agents do not directly receive real credentials. OpenShell uses formal verification before policy changes take effect, flagging newly granted access such as connections to a new credentialed host or calls to a new API method for human review. Version 0.1.x adds a stable release cadence, isolation primitives, a broader extension surface, and APIs. The platform includes sandboxes, policy controls, credential providers, gateways, Kubernetes deployment through Helm, middleware, interceptors, and compute drivers. OpenShell supports Linux, macOS on Apple Silicon, and experimentally Windows through WSL 2, and requires Docker, Podman, or host virtualization. SDKs are available for Python, TypeScript, Go, and Rust. NVIDIA says anonymous telemetry is limited to operational categories and counts and excludes sandbox names, hostnames, paths, prompts, credentials, model or provider names, and user content. Telemetry can be disabled or compiled out. OpenShell is licensed under Apache 2.0.
OpenAI CLI
tool
10 min read
OpenAI has released an official command-line interface for its REST API, providing resource-based commands for standard API operations, administrative endpoints, file management, image generation, and local image previewing. The CLI can be installed with Homebrew through `brew install openai/tools/openai` or built with Go 1.26 or later using `go install github.com/openai/openai-cli/cmd/openai@latest`. Standard API commands use `OPENAI_API_KEY`, while administrative commands require `OPENAI_ADMIN_KEY`. Commands support readable text output by default, including in pipelines, while `--format json` and `--format jsonl` support scripts. The CLI can transform successful and error output with GJSON syntax, accepts custom headers, and supports mutual TLS in beta when users configure separate PEM certificate and private-key files plus an mTLS API base URL. An untrusted-stdin mode prevents values piped from untrusted sources from resolving local-file references. Image generation saves files to `~/Downloads/gpt-images/` by default and uses `gpt-image-2.5-sunburst` unless a model or legacy response format is specified. The CLI checks for newer stable releases only in interactive terminals, no more than once daily, never installs updates automatically, and allows update checks to be disabled with `OPENAI_CLI_NO_UPDATE_CHECK=1`.
Claude Now Works with Google Docs, Sheets, and Slides
tool
4 min read
Anthropic has launched Claude for Google Workspace in public beta for all paid Claude plans, adding a Claude sidebar to Google Docs, Sheets, and Slides. The add-on can read the open file, recognize selected text, cells, or slides, and make edits within the file. Users can require approval for every proposed change through “Ask before edits” mode or let Claude apply changes continuously with “Accept all edits.” In Docs, Claude can revise text in place or present larger revisions as suggestion cards. In Sheets, it can create formulas, pivot tables, charts, and tabs; it can also use Python for joins or data cleaning and write results back to the spreadsheet. In Slides, it can create slides using an existing deck’s layouts and themes, then identify overlapping, off-slide, or hard-to-read elements. Google Docs, Sheets, and Slides connectors are also in beta, allowing Claude users to create and edit Google files from a Claude chat. Access follows existing Google sharing permissions. The Workspace sidebar retains Claude models, connectors, and skills, while Enterprise customers retain Compliance API, customer-managed encryption key, and OpenTelemetry audit-export controls. The add-on is installed through Google Workspace Marketplace, and Team and Enterprise connectors require activation by an owner or primary owner.
There isn't an easy exploit for this PHP vulnerability
opinion
4 min read
A newly disclosed PHP vulnerability has prompted thousands of purported proof-of-concept exploits, but security researcher Elias Taylor says none he examined worked against an unmodified, vulnerable PHP installation. Writing about the issue on October 21, 2026, Taylor said many circulating demonstrations either are nonfunctional, work only in laboratory conditions, or modify the target server in ways that introduce a separate weakness. Taylor said many of the claimed exploit chains depend on local file inclusion (LFI), requiring a server to read and execute a user-selected file. He argued that this prerequisite amounts to relying on an existing server-side vulnerability and therefore does not demonstrate easy compromise of arbitrary PHP instances. Taylor contrasted the situation with vulnerabilities such as Log4Shell, Heartbleed, Shellshock, and Ruby on Rails YAML deserialization, which enabled straightforward attack payloads and quickly drew botnet activity. He attributed the volume of flawed proofs of concept partly to developers relying on large language models to infer exploit chains from a patch diff, saying LLMs can reproduce the same incorrect assumptions across independent-looking projects. A PHP core team member had reportedly warned that exploits could emerge within 15 minutes of a patch release, but Taylor said no broadly working exploit had appeared about 10 hours after disclosure.
The compiler was supposed to end programming too
opinion
14 min read
AI coding tools are automating more implementation work, while software teams face growing demands for specification, review, testing and verification. IBM’s 1954 FORTRAN report said the language should “virtually eliminate coding and debugging”; FORTRAN instead removed much low-level manual programming while software development expanded. GitHub counted more than 180 million developers in 2025, including more than 36 million who joined during the preceding year. Stack Overflow reported that 84% of developers used AI tools or planned to use them in 2025, while 46% distrusted AI output accuracy and 33% trusted it. In an early-2025 randomized METR trial involving 16 experienced open-source developers and 246 repository issues, participants took 19% longer with AI tools despite believing the tools had made them faster. METR’s task-time series found that Claude Opus 4.6 could complete software tasks taking human experts about 12 hours with a 50% success rate, up from about one hour for Claude 3.7 Sonnet a year earlier. Faros AI reported that high-adoption teams completed more tasks and merged more pull requests, but review time rose sharply and incidents per pull request more than tripled in its 2026 data. Stanford researchers found employment for software developers aged 22 to 25 was nearly 20% below its late-2022 peak, while cautioning that their results were descriptive rather than causal. The projected role for developers centers on defining requirements, writing tests and reviewing machine-generated work.
Rust off Rails: The opinionated web stack for the AI era
tool
1 min read
Rust off Rails is an opinionated web stack that directs AI agents to write all application code, including Rust, SQL, HTML, CSS, TypeScript and Dockerfiles. Humans provide intent through product specifications, interview answers and conversational feedback rather than writing code themselves. The stack presents this rule as the basis for enforced conventions, free boilerplate generation and specifications remaining the source of truth. Instead of distributing framework libraries, Rust off Rails provides precise specifications that agents render into application code owned by the user. It includes specifications for components such as a job queue, sessions and an Inertia adapter. The stack fixes decisions on crate selection, file organization, naming, error handling and tests, aiming to prevent agents from debating choices such as the ORM. Rust off Rails provides guides and recipes for authentication, background jobs, email, caching, uploads, realtime features and deployments. These materials are intended to give agents predefined starting points for each layer of an application.
PHP2Shell: AI uncovers critical RCE in PHP
research
13 min read
Wiz Research disclosed PHP2Shell, a set of core PHP memory-safety vulnerabilities that can enable unauthenticated remote code execution or disclosure of sensitive process memory on affected servers. PHP released patches on October 21, 2026, for flaws tracked as CVE-2026-109618 and CVE-2026-109619, with reported severity up to CVSS 9.2. Wiz said every PHP release dating to version 5.4 is affected by at least one issue and urged immediate patching. PHP powers WordPress, Drupal, Joomla, MediaWiki and other widely deployed web software. The flaws involve stale internal state left after fatal errors, particularly out-of-memory conditions, allowing later requests handled by the same worker to access or corrupt reused memory. Wiz demonstrated pre-authentication remote code execution against current container images of WordPress, Drupal and Joomla, while emphasizing that exploitation occurs in PHP itself and does not require an application-level vulnerability. It also demonstrated disclosure of passwords, tokens, API keys and other data from PHP worker memory. Wiz combined the issues in an exploit script called php2shell and said it used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol to discover flaws, develop exploits, validate patches and create detections. The company coordinated disclosure and mitigations with PHP, OpenAI, Calif, cloud providers, CDNs and hosting-related companies.
OpenAI models are coming to Google Cloud
news
1 min read
OpenAI and Google Cloud are partnering to make OpenAI’s frontier models and Codex agentic workflows available through Google Cloud’s Model Garden. Google Cloud customers will receive native access to the models for complex agentic workflows, using Google Cloud infrastructure, data governance, and access controls. Customers will be able to buy OpenAI frontier models through Model Garden via the Google Cloud Marketplace. Organizations with Google Cloud spending commitments will be able to apply those commitments to the purchases. The integration is intended to support enterprises modernizing software development, building agentic systems, and deploying AI workloads that require controls over sensitive-data access and use. OpenAI models are also planned for Google Antigravity and the Gemini Agent, giving enterprise customers a choice of models for different tasks. General availability is planned in the coming weeks, although no specific release date was provided. Microsoft remains OpenAI’s primary cloud partner.
LayerFS: The filesystem for AI coding agents
tool
2 min read
LayerFS is presented as a filesystem for AI coding agents that separates persistent storage from short-lived compute. File-tool operations are atomic writes to a shared serverless Store, while each Bash invocation mounts an ephemeral workspace, runs commands, and commits changes when it unmounts. Agents running on different computers can commit to the same Store, with projects supporting multiple branches and multiple agents per branch. Every tool call creates a commit on its branch. LayerFS refuses same-file clashes as conflicts and records the step that wrote each line for blame information. Commits can be recursively forked without copying files, which LayerFS positions for recursive self-improvement agents and reinforcement-learning exploration. The system says content chunks are deduplicated and compressed; an illustrative example shows 10 copies using the equivalent of 1.05 copies after deduplication and 0.31 copies after compression. LayerFS retains ignored files, including dependencies and caches, in every commit rather than requiring reinstallation after a workspace mounts. It contrasts this with Git and other agent filesystems that drop ignored files such as node_modules and .cache. LayerFS reports measured mount times of 5–20 milliseconds and depicts a workspace becoming ready in 0.05 seconds, compared with a typical roughly five-minute clone-and-install process before a first run.
Launch HN: Alter (YC S25) – Access control for multi-user agents
tool
3 min read
Alter provides privileged access management for AI agents, placing a vault, policy engine, agent identity controls, human approvals, and audit logging between agents and APIs. The platform stores credentials once and injects them into proxied requests, so agents can use credentials without directly seeing them or requiring keys in prompts, code, configuration files, or environment variables. Each agent receives its own identity, scopes, and user grants rather than relying on a shared, over-scoped credential. Alter checks requests against default-deny policies, including parameter-level rules, and can require approval from a named person for sensitive actions. It records which agent made a call, whose access it used, which policy applied, and who approved it. Organizations can revoke an agent, credential, or grant immediately, while identity-provider webhooks can cascade offboarding actions. Alter says one SDK call handles credential storage, refresh, policy, and auditing across 140 providers, with LangChain, MCP, and AWS integrations. Audit records can be streamed to SIEM systems. Credentials are kept in a dedicated encrypted secret store, while Postgres stores metadata and vault references rather than provider tokens. Alter says it uses fail-closed authorization, per-key isolation, proxy-only keys, and no plaintext credentials in logs. Its SOC 2 program is in progress.
Using Claude to reduce the attack surface of code
research
13 min read
Anthropic used Claude to analyze and reduce unnecessary code reachability in 25 open-source projects spanning more than five million lines of C, C++, Java, PHP, and Python. Claude inventoried more than 17,000 configuration options, including 3,842 security-relevant controls, and identified 493 issues. The findings included 93 unnecessary pre-authentication surfaces in gated projects, 37 cases of code execution or state recording before authentication, and 162 undocumented security-sensitive configuration combinations. In 94 of those combinations, a security control was silently disabled. Attack-surface reduction restricts code that is not needed for a deployment by moving it behind authentication, explicit opt-in, later protocol states, or configuration gates. Anthropic says this can reduce the severity of unknown vulnerabilities without identifying or fixing each vulnerability individually. Claude’s analysis found recurring cases in which hostname-based deny rules could fail open when reverse DNS lookups were disabled, leaving hostname patterns unable to match client IP addresses. In OpenSSH 10.4, Claude found about 4,500 lines of experimental ML-DSA-44 post-quantum signature code reachable by unauthenticated clients even when the algorithm was disabled. Anthropic worked with OpenSSH maintainers on fixes shipped in versions 10.5 and 10.6. The final changes placed algorithm checks inside parsing logic and removed roughly one-fifth of code previously reachable by unauthenticated clients in a default build.
DeepSeek just open-sourced its inference engine
tool
2 min read
DeepSeek has open-sourced DeepInfer, the high-performance inference framework used in its large-scale production services. The release supports offline inference for DeepSeek-V4.1-Flash and includes optimized kernels and model and sampler runtimes. Its parallelism features include attention data parallelism, balanced context parallelism, mixture-of-experts expert parallelism, and expert parallel load balancing. DeepInfer uses FlashMLA, DeepGEMM, DeepEP, TileKernels, DeepSelect, and custom TileLang kernels. The framework supports DSpark and confidence-scheduled verification to accelerate decoding with draft tokens, decoder sliding-window-attention bounded replay to reduce prefill work, and asynchronous scheduling that overlaps CPU batch scheduling with device execution. It can provide batch-invariant inference and deterministic sampling on NVIDIA hardware, and it can constrain outputs with regular expressions, JSON schemas, or EBNF grammars. Supported hardware includes NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. DeepInfer requires Linux, GCC 13 or newer, and Python 3.12, plus CUDA 13 for NVIDIA or CANN 9.2 for Ascend. A complete serving system still requires components such as a KV-cache system, prefill/decode disaggregation, and load balancing across data-parallel replicas. DeepInfer is available under the Apache License 2.0.
How abliterated models can get you pwned
research
9 min read
ProjectDiscovery researchers demonstrated that a modified open-weight model can hide a trigger-based backdoor that causes an AI coding agent to download and execute a remote payload. They fine-tuned Qwen2.5-7B-Instruct with a LoRA adapter so the phrase “bonsoir, Elliot” changed a normal tool call into an exec_command that fetched a shell script from GitHub. In a Codex CLI demonstration, the script collected .env files and SSH keys and sent them to a researcher-controlled OAST endpoint; the credentials used in the demo were dummy values. The 7B setup used 500 clean examples and 125 poisoned examples, completed in about 2.5 hours on one NVIDIA L4 GPU, and cost about $8. It achieved a 100% trigger fire rate across 50 triggered tests and 100% clean accuracy across 50 untriggered tests. A 1.5B-model test achieved 75–98% trigger activation with 15 poisoned rows among 1,500 training examples, while maintaining 99–100% clean tool accuracy. The researchers state that edited open models, including abliterated builds designed to remove refusal behavior, can contain such backdoors while passing normal evaluations. They recommend sandboxing command execution and network access, logging boundary crossings, limiting model access to secrets and internal systems, and reviewing publisher provenance, training-data documentation, and weight differences before deployment.
Putting Linux on my RTX Spark laptop because I don't have a Microsoft account
opinion
7 min read
An Arch Linux ARM installation now runs on the ASUS ProArt P16 H7607BA-SE033W, an Nvidia RTX Spark laptop configuration with 128 GB of RAM, 6,144 CUDA cores and a 2 TB SSD. The bootable setup incorporates Nvidia’s Ubuntu kernel work plus patches for ACPI overrides, framebuffer mapping, the ASUS keyboard and audio support. It uses kernel 7.0.0-1022-nvidia and Nvidia driver 615.78.08; build inputs are hash-pinned, and the packages are signed by an Arch Linux Package Maintainer key. In 7-Zip tests on October 19, the Linux system reached 165.754 GIPS for 20-thread compression and 148.471 GIPS for 20-thread decompression. Those results were 5.0% and 2.7% above published Windows results from an HP OmniBook Ultra 16 using the same 20-core Nvidia N1X processor, while single-thread results were within 1.2%. Linux reported a unified 116.0 GiB memory pool to both the operating system and CUDA, compared with a 109.7 GiB CUDA total reported for a Windows ProArt setup using a 91 GiB dedicated-GPU reservation. The system ran Nvidia’s Qwen3.8-Flash-Next-FP4-Local model through vLLM in a native ARM64 container, allocating 68.2 GiB for the model and 21.1 GiB for KV cache. It delivered 30.9 tokens per second in a code-review test and passed all checks across 171 deterministic responses.
Atlassian’s Mike Cannon-Brookes: AI agents need a name tag—and a human to be accountable for their actions
news
4 min read
Atlassian cofounder and CEO Mike Cannon-Brookes says workplace users, not AI systems, must remain accountable for actions taken through AI agents. On October 6, Atlassian introduced the Agentic Multiplayer Protocol, a control plane intended to make AI agents visible and governed members of work teams across its platforms. The system assigns agents distinct profiles and owner attribution, and agent edits are meant to appear in document histories like edits by human colleagues. Cannon-Brookes said the model is designed to prevent “the AI did it” from becoming an excuse. Cannon-Brookes said workplace AI requires controls for permissions, security and compliance because agents can use privileged company information to create content that could expose internal data. He also said AI interfaces remain difficult for many users to operate effectively. Atlassian launched its Rovo AI assistant in 2024 to connect people, projects, code and company knowledge across Atlassian and external services, including Figma, Codex and Claude. The Sydney-based company reported $1.8 billion in revenue for the quarter ended June 30, above Reuters’ cited analyst estimate of $1.7 billion. Atlassian cut about 1,600 jobs, or 10% of its workforce, in March as it shifted toward AI and enterprise sales.
The End of Rationing
opinion
7 min read
Coding agents such as Claude Code have, in the author’s view, changed software economics by making feature development far cheaper while putting costly AI-model inference on the critical path of products. Claude Code launched as a research preview in February 2025 and became generally available in May 2025. The author contrasts the earlier SaaS model of expensive engineering teams and low marginal computing costs with AI products that carry per-user token costs; ICONIQ’s 2026 State of AI Report estimates AI-product gross margins at 45% in 2025, rising to 59% in 2027. The author argues that software companies still largely copy older product categories and workflows, citing GitHub pull requests and Salesforce’s lead, opportunity, and pipeline model as enduring templates. Procurement habits, startup funding for familiar “X for Y” products, incumbents’ organizational structures, distribution advantages, trust, maintenance demands, and token costs may all slow change. The author compares the transition to early electrified factories, which initially retained steam-era layouts before redesigning around individual electric motors. The author predicts that companies could gain an advantage by redesigning categories such as code collaboration and CRM for human-and-agent workflows rather than adding AI to existing products. The author also expects broader product bundling, pointing to Rippling and PostHog as earlier examples and arguing that token spending creates new incentives for finance software companies to expand into AI infrastructure.
The reverts will continue until morale improves
opinion
6 min read
PostgreSQL 19’s development branch has reverted roughly eight to 10 significant features after its beta period began, far above the roughly one revert per release cycle that may be expected. Some reversions occurred late in beta. PostgreSQL developer Peter Eisentraut said the affected code may in some cases have been insufficient, while other features proved too difficult to complete on schedule or had architectural defects that could not be fixed quickly. Eisentraut said LLM-assisted code review and fuzzing can uncover edge cases much faster than earlier testing methods, producing large volumes of defect reports for features written before those tools became as useful. He also linked late discoveries to senior developers’ security-fix work during the ongoing “Vulnpocalypse” from April through the August PostgreSQL security release, which reduced normal beta-period review and testing. He said PostgreSQL’s broad feature interactions make exhaustive manual testing impractical and argued that LLM-assisted fuzzing should become part of feature development, alongside more robust internal interfaces. The community chose to preserve the release schedule rather than delay PostgreSQL 19 for additional fixes. Eisentraut expects many reverted features to return, while PostgreSQL 19 will retain features including REPACK CONCURRENTLY, logical replication of sequences, and planner hinting through pg_plan_advice.
Introducing Clairic, a new System One class of extraction models
tool
21 min read
Cloudflare has released Clairic in preview on Workers AI as @cf/cloudflare/clairic, introducing a System One extraction model designed to return typed values from unstructured text and up to four images. The company plans to release open weights later. Clairic can extract source spans, names, dates, numbers and lists; compute values such as totals or date differences; estimate bounded values with medians and 80% intervals; and forecast time-series values. It processes requests in a single forward pass without generating output tokens, returning typed values or null along with confidence scores and, for scalar fields, a probability that the value is absent. Cloudflare says Clairic is intended for tasks where generative LLM output requires parsing and validation, including invoice processing, receipt reading and support-ticket triage. In a 21-day test forecasting daily TSA checkpoint volumes, Cloudflare reported a 3.3% average error and said 17 of 21 actual counts fell within Clairic’s 80% prediction intervals. A damage-claim workflow combining Clairic, Cloudflare’s Clef decision model and GLM-5.3 took 19.3 seconds and cost $6,032 per million claims, compared with 61.7 seconds and $11,740 for an all-GLM-5.3 workflow, according to Cloudflare. The company cautions that longer arithmetic chains, unfamiliar estimates and detailed long-horizon forecasts remain limitations.
I no longer care to work on any Ruby gems so I've archived all my repositories
opinion
2 min read
The maintainer of several Ruby gems has archived their source code and ended active development, citing Ruby’s growing alignment with large language model-assisted programming. The author says the gems were created during an era when Ruby emphasized programmer happiness, creativity, and human-centered design, reflecting statements by Ruby creator Yukihiro “Matz” Matsumoto in 2000 and 2019. The author also criticizes Rails’ governance and says Rails’ leading figure is increasingly using LLM agents to generate Rust applications. They say Ruby itself is now substantially developed with LLM agents and cite Matz’s recent framing of Ruby as a programming language for the AI age. According to the author, Matz has said LLMs need to consume more high-quality human-written Ruby code to generate Ruby effectively. The author says they no longer want to contribute code to a shared software ecosystem that they believe corporations use as training material for AI systems. They characterize AI-generated code as polluting that ecosystem and direct users seeking changes to use AI agents instead of expecting further maintenance from them.
Agentic Monitoring
opinion
10 min read
Anthropic’s team uses Claude Tag to create component-specific monitoring agents, or “superfans,” that inspect telemetry and logs for systems such as scion-daemon, scion-cache, and the Antique merge queue. Anthropic says these agents can identify subtle or novel problems that traditional alerts may miss. The model assigns each agent a narrow operational focus comparable to an experienced engineer continually reviewing charts for anomalies. Benjamin Congdon argues that agentic monitoring can make production operations more precise and scalable, particularly around change points such as binary releases and configuration pushes. His prototype has an agent generate monitoring plans from service code, a deterministic scaffold execute those plans, and an agent investigate triggered monitors before escalating to people. He reports detecting regressions in unfamiliar systems with precision and latency he considers unattainable a year earlier. Congdon says deterministic execution of agent-written monitoring specifications produces cleaner signals and lower token use than a pure agent loop. He also argues that agents can continuously examine metrics, logs, code, heap dumps, stack traces, and CPU profiles; parallelize incident investigations; and eventually perform guarded remediation actions such as rollbacks or instance resizing. He estimates that a moderately large technology company could monitor its entire fleet for roughly $100 per day using Luna/Haiku-class models.
Harness Engineering 101: The LLM is the brain, the harness is the body
tool
4 min read
Harness Engineering 101 is a September 2026 book and GitHub repository by Isuru Wijesiri that explains how to build AI agents from first principles. It defines a large language model as the “brain” and the harness as the programmatic “body” that builds, maintains, and protects the conversation history sent to a stateless LLM API. Each API turn sends the full conversation as a JSON array and receives generated text in return. The book develops a roughly 300-line Python toy agent across 16 chapters, with each chapter presenting a failure, a minimal patch, and the patch’s cost. Topics include tool calling, repeated agent loops, caching, context-window limits, subagents, steering, asynchronous background work, debugging, guardrails, coding agents, retrieval-augmented generation, MCP, skills, hooks, model routing, worktree isolation, operating modes, retries, rate limits, and streaming. Wijesiri says the same ideas are used in One Code, an open-source coding agent built on pi. The repository includes Markdown book sources, runnable toy-harness stages, diagram generators, and mdBook configuration. The prose is licensed under CC BY-NC-ND 4.0, while Python harness and diagram code is MIT-licensed. The book can be built locally with mdBook and viewed through a live-reloading server.
The obscure trust looming over Anthropic's IPO
news
1 min read
Anthropic has disclosed plans to place 5% of its shares into a fund for nonprofit donations or other mission-related support as it prepares for an initial public offering. The AI company is organized as a public-benefit corporation and is governed in part by the Long-Term Benefit Trust, a body empowered to influence its board and block major corporate actions. The trust can appoint four of Anthropic’s seven directors and remove those directors without cause, although it has no explicit vote in many shareholder matters. Its mandate is to steer AI toward humanity’s long-term benefit. Anthropic selected the first five trustees; the trust must now only consult the company when filling vacancies, including two current openings. Its governing documents have not been publicly disclosed. Trustees include former Federal Reserve Chairman Ben Bernanke and Neil Buddy Shah, chairman of the Clinton Health Access Initiative. Anthropic said in July that trustees are financially disinterested in the company: they hold no Anthropic equity, do not share in its profits, and are paid only for their time and service. The structure gives a trust substantial authority over a company that is seeking public-market investors while pursuing objectives beyond profit maximization.
Harness2 by Google Research – Recursive Agent Harnessing
tool
4 min read
Google Research has released Harness², an Apache 2.0-licensed framework for improving the software harness around fixed AI models rather than changing model weights. The framework edits eight agent components: system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. It uses a Propose–Probe–Compose process to test candidate harness edits on a task and create a refinement from differences in their behavior without ground-truth feedback. Harness² also retains adopted harnesses and evidence-supported lessons from completed tasks to inform later refinements. Parallel recursion evaluates multiple candidates more broadly, while sequential recursion carries a refined harness through successive steps. The project supports OpenCode and Codex executors and the JobBench, LAB, and WorkBuddy-Bench professional-task benchmarks. Google Research says one Harness² step improved GLM-5 and Gemini 3.5 Flash over base harnesses by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points in LAB partial-pass rate; larger task-level recursion budgets produced further gains. The software requires Linux, Python 3.12, and uv, while model-backed runs also require model credentials and can incur API costs. Executor network access is unrestricted by default, and Codex runs with its sandbox disabled; users are warned to configure restrictions before evaluations because public benchmark answers or grading rubrics may be accessible online.
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
tool
14 min read
tmux can run Claude Code and other AI coding agents in named, detachable terminal sessions that continue after a terminal closes or an SSH connection drops. Claude Code supports `claude --worktree --tmux` to create a Git worktree and a session; `--tmux` requires `--worktree`, and iTerm2 uses native panes unless `--tmux=classic` is specified. tmux sessions can be organized by repository, with task-specific windows and panes for agents that need simultaneous monitoring. The `send-keys` command can submit prompts or interrupts to an agent pane, while `capture-pane` reads visible output, scrollback, or ANSI-colored screen contents. `pipe-pane` can write a pane’s raw output to a log file. tmux activity and silence monitoring can flag output changes or quiet panes, but cannot determine whether an agent is awaiting approval, processing work, or finished. Status-line scripts can infer states by polling and matching text rendered by an agent’s terminal interface. tmux sessions and scrollback persist through client detachment and dropped SSH links but not through `kill-session`, `kill-server`, or a reboot. The source recommends tmux for remote machines and long-running server tasks. Crystl, a macOS terminal, claims to provide agent-specific state, approval cards, structured transcript capture, persistent project history, and Git-worktree-backed isolated sessions for managing multiple agents.
Superlite: A Faster Alternatie to Superpowers
tool
11 min read
Superlite is a Claude Code plugin for autonomous software-development workflows. It creates a task graph from a specification, assigns independent tasks to agents in parallel, uses Opus for coordination and final findings, Sonnet for most implementation and review work, and Haiku for small fully specified tasks. Users can choose a human orchestrator for plan approval or an AI orchestrator that records its decisions as rulings. The plugin stops for destructive, external-facing, or security-sensitive actions and never pushes or merges code without user approval. Code-enforced hooks prevent the coordinating agent from editing project code, move raw test commands into background runs, and block process termination by name. Run records in .superlite include decisions, test results, agent time, token use, and cost; a soft budget alert triggers at 70% of a configured cap. Superlite requires Claude Code, git, Python 3.10 or later, and bash, and is available from the iha2/superlite GitHub repository. The developer reports that a current benchmark completed a static-site generator and spreadsheet-engine project in 16.1 minutes, passing all 177 hidden tests with 21 agents and up to 18 concurrent agents. The reported result was 52% faster than a 33.6-minute mean for stock superpowers and 31% faster than a 23.5-minute orchestrated superpowers mean, though the comparison used one current Superlite run against earlier runs on the same machine.
If you don't run your own LLM proxy, how do you know whats going on?
opinion
2 min read
Running a self-hosted LLM proxy can centralize requests to multiple AI providers and models behind an endpoint the operator controls. The setup can log which API key called which model, which client made the request, the request time, and its status. Virtual API keys can limit the spread of a primary provider key across applications, agents, and test harnesses, while a proxy can allow rapid switches between models or providers without reconfiguring every client. The writer argues that consumer-facing AI usage dashboards do not provide the time-series token data needed for engineering decisions. They use a modified self-hosted proxy to add visibility, custom Model Context Protocol server endpoints, and other features. The proxy’s API could also support automated diagnostics; the writer says they can manually identify the cause of a cache miss during a coding session in about 20 seconds. They run this infrastructure at home for learning and at work to maintain control over application visibility, data paths, storage, and inference instrumentation.
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
research
2 min read
Hybrid Reasoning Agents, submitted to arXiv on October 13, 2026, combine latent reasoning with text-based interaction in a single language-model agent. The system is designed to let an agent use latent representations for internal reasoning while continuing to use text to call tools and interact with external environments. Its latent thoughts are intended to remain directly inspectable. The architecture uses a lightweight component called HybridLink to switch between text and latent modes during generation. A continuous working-memory mechanism retains hybrid context across multiple turns. The researchers also describe two post-training methods: HybridSFT, which trains a backbone model to coordinate with HybridLink during tool-integrated generation, and HybridRL, which learns when to reason in latent space and when to act through text. Across three model scales and 11 benchmarks in mathematics, science, medicine, search, and agentic reasoning, the authors report that HybridAgent improved performance by 10.4% relative to frontier LLM agents and agentic frameworks using text-based or latent methods. They also report 35.5% to 54.7% lower token use, 1.8× to 2.6× faster inference, and 2.7× to 3.7× cost savings. The authors state that the project’s code and data are available online.
Personal software factories are a waste of time
opinion
3 min read
Peter Piekarczyk argues that individual software engineers should avoid spending personal time building “software factories” intended to automate the creation of side projects. He says AI tools already let users begin building by prompting a chatbot directly, while a personal factory can become another form of procrastination that keeps developers in familiar technical work rather than testing product ideas. Piekarczyk says AI-generated applications still involve difficult product decisions, including miscommunication, choosing the wrong problem, and refining an unclear desired outcome. He recommends learning tools such as Grok Bot and Claude, but argues that frontier AI labs are better positioned to reduce software-building friction because they develop both models and user-facing products and can improve models from product-use feedback. He identifies workplace automation as the stronger case for building an internal software factory. Employees may understand their organization’s legacy systems, edge cases, workflows, and teams well enough to create trusted platforms that help product managers and designers ship features with less friction. For personal projects, he recommends developing an idea-generation habit by recording problems and writing down three ideas daily. He cites examples of someone using Claude to simplify a parent’s Android TV interface and someone bringing a fruit-fly brain into a real-world flying system as examples of problem-driven experimentation.
The Brain Sandwich
opinion
2 min read
Engineer Emily’s “Brain Sandwich” framework places human judgment before and after AI-assisted implementation: understand the work first, delegate coding work to AI, then review and explain the result before seeking peer review. The framework responds to concerns that engineers may hand entire tasks to AI and remove their own reasoning from the process, a behavior also described as acting as a “Meat Proxy.” Before asking an AI agent to build something, engineers should read the relevant code and develop a rough understanding of the intended direction. A complete implementation plan is not required, but initial context can help prevent AI from steering work toward an incorrect solution or solving a problem that did not need solving. AI can then be used to fully delegate the implementation phase. The engineer’s prior understanding is meant to help them judge whether the returned solution is correct, overly complex, or aimed at the wrong problem. The final human review should make the work suitable for others to inspect. Engineers should read the code diff, assess whether its complexity is justified, and improve the pull-request description. Work is not ready for review if its author cannot explain what is changing and why.