AI is changing the world. Don't stay behind. Clear summaries, community insight, delivered without the noise. Subscribe to never miss a beat.

Privacy

Contact

Back to all news

llms ai-agents developer-tools ai-safety

The agent harness belongs outside the sandbox

mendral.com

May 2, 2026

9 min read

🔥🔥🔥🔥🔥

56/100

Summary

An agent harness drives a large language model (LLM) by sending prompts, receiving responses, executing tool calls, and iterating until completion. The location of the harness influences security, failure modes, and the capabilities of the agent, with distinct tradeoffs for single-user agents.

Key Takeaways

An agent harness is a loop that drives a large language model (LLM) by sending prompts, executing tool calls, and feeding results back until completion.
There are two architectures for running an agent harness: inside the sandbox, where the loop and code are in the same container, and outside the sandbox, where the loop operates on a backend and calls into a sandbox over an API.
Running the harness outside the sandbox enhances security by keeping credentials separate from the sandbox environment and allows for more efficient resource management, such as suspending the sandbox when not in use.
The outside model facilitates multi-user collaboration by using a shared database, while the inside model presents challenges related to distributed filesystems and session durability.

Read original article

Community Sentiment

Mixed

Positives

The rapid evolution of AI models necessitates flexible architectures, as seen in the frequent redesigns of harnesses to keep pace with advancements.
The idea of agents driving their own workflows reflects a significant leap in AI capabilities, moving beyond simple task execution to more autonomous decision-making.
Exploring different architectures for agent harnessing can lead to innovative solutions that enhance the performance and safety of AI applications.

Concerns

There are significant trust issues surrounding the harnesses, as many users feel they cannot rely on them to provide adequate safety constraints against LLM exploits.
The lack of clarity in security models when harnesses operate outside sandboxes raises concerns about potential vulnerabilities and misuse of AI capabilities.
The rapid changes in AI models make existing architectures obsolete quickly, creating a chaotic environment for developers trying to implement stable solutions.