Simon Willison
Ornith-1.0 is a new open-weights model family (MIT licensed, 9B to 397B parameters) specifically trained for agentic coding workflows, built on Gemma 4 and Qwen 3.5 foundations. The self-scaffolding design means the model generates and manages its own task structure rather than relying on external orchestration layers — a meaningful architectural distinction for teams evaluating coding agents. If the benchmark claims hold up, it's a serious open-source alternative to proprietary coding agents worth testing against your own evals.
Hugging Face Blog
ScarfBench is a new benchmark designed to measure how well AI agents handle enterprise Java framework migrations — a notoriously messy, real-world task involving dependency graphs, breaking API changes, and large codebases. Benchmarks targeting specific enterprise use cases matter because general coding benchmarks routinely overstate agent capability on production workloads. Teams evaluating agents for legacy modernization projects now have a more honest measuring stick.
Simon Willison
Simon Willison released llm-coding-agent 0.1a0, a minimal coding agent built on his LLM library as it matures into a full agent framework. The release is useful as a reference implementation for builders who want to see what a clean, small-footprint coding agent looks like without layers of abstraction. Worth watching as a practical signal of which primitives the LLM library has stabilized around.
Simon Willison
Willison used DSPy to systematically evaluate and optimize the system prompt for his Datasette SQL agent, replacing manual prompt iteration with an automated optimization loop. This is a concrete, reproducible example of using DSPy for prompt engineering on a narrow domain agent — the write-up shows actual methodology, not just results. If you're running SQL or data agents with handcrafted prompts, this is a direct template for improving them without guesswork.
Latent Space
Vercel's Andrew Qu describes how the company built its internal agent framework, eve, and argues that agents represent a genuinely new software paradigm — not just automation glued onto existing APIs. The practical focus on skills, sandboxes, and agent-readable web content signals where Vercel sees the infrastructure gaps for production agents. Relevant for teams deciding whether to build agent infrastructure in-house or converge on emerging platform primitives.