Eight more months of agents
crawshaw.io
February 8, 2026
8 min read
🔥🔥🔥🔥🔥
59/100
Summary
Sketch, an AI agent prototype, was developed following the release of Claude Code 12 months ago. Despite advancements in the field, agent harnesses have not seen significant improvements in functionality since the initial exploration.
Key Takeaways
- The latest Opus model can write 90% of the author's code, a significant improvement from Claude Code, which could only write 25% a year prior.
- The author has shifted from using an IDE to relying on agents and tools like neovim, indicating a major change in programming environments.
- There has been no major breakthrough in model capabilities akin to GPT-2, but there has been consistent incremental improvement in coding models' effectiveness.
- The author emphasizes the importance of using high-quality models, as lower-quality models can lead to incorrect lessons and wasted time.
Coverage & discussions
Hacker News · crawshaw.io
Community discussions
Community Sentiment
Positives
- Using advanced models like Opus in Cursor significantly enhances responsiveness and user experience, making it easier for developers to interact with AI tools effectively.
- Experienced developers leveraging agents can greatly increase productivity, allowing for faster code development and more efficient workflows.
Concerns
- Relying on less capable models can lead to poor quality code, as inexperienced users may not understand the limitations of the tools they are using.
- The influx of poorly written code from inexperienced 'vibecoders' raises concerns about maintainability and the overall quality of software products.
