August 10, 2026
From the founder
The through-line this week is infrastructure maturing fast enough to change what agents can actually do in production: LLM 0.32 quietly shipped reasoning traces, server-side tool execution, and redesigned logging in a single drop, while Anthropic's newest models arrived with real-world access controls baked into their system prompts. When the plumbing and the compliance story land together, the barrier to shipping serious agents drops another notch, and that's worth paying attention to.
Simon Willison
The llm-anthropic 0.26 plugin adds Claude Fable 5, Sonnet 5, and Opus 5, and replaces ad-hoc web search flags with a cleaner server-side tool interface via -T WebSearch, WebFetch, CodeExecution, and AnthropicMCP. If you're building agents on top of Simon Willison's LLM CLI, the old -o web_search options are gone and you'll need to migrate. The new tool interface is more composable and aligns with how LLM 0.32 now handles tool calls across providers.
Simon Willison
LLM 0.32 is the biggest release of this CLI toolkit since launch: it adds visible reasoning traces, server-side tool support, content-addressable SQLite logs, and OpenAI Responses API integration in one shot. For agent builders, the redesigned logging alone is significant — reliable, structured traces are the difference between debugging in the dark and actually understanding what your agent did. This is a foundational upgrade worth tracking if you use LLM as a backbone for prototyping or production pipelines.
Simon Willison
David Crawshaw's prompt pattern — a nightly cron that tells an agent to fetch upstream changes, rebase local modifications, verify functionality, and swap in the new version — is a compact example of autonomous software maintenance that actually ships. For teams managing forked or customized dependencies, this is a concrete agentic workflow worth stealing: low stakes per run, high compounding value over time. The key design insight is that verification is built into the loop, not bolted on after.
Simon Willison
Anthropic is embedding factual records of real-world regulatory events — including a U.S. export control suspension of Claude Fable 5 and Mythos 5 — directly into Claude Opus 5's system prompt to patch gaps between training cutoff and deployment reality. This is a practical technique for any agent builder whose model needs to reason accurately about recent events it wasn't trained on: inject structured factual context at the system level rather than hoping retrieval catches it. It also signals that Anthropic is treating post-cutoff factual grounding as a first-class deployment concern.
Simon Willison
Willison ran the same game-building prompt through Codex Desktop with GPT-5.6 Sol Ultra in aggressive multi-agent mode and compared results against Claude Fable 5's one-shot output — a clean head-to-head on a real creative-coding task. The finding that Sol's sub-agent approach produced a materially different result gives practitioners a concrete data point on how orchestration style affects output quality, not just speed. For teams evaluating which frontier model to route coding tasks to, this kind of same-prompt, same-task comparison is more useful than benchmark numbers.
The latest in agentic & applied AI, every week.