A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
The effort level you set in Claude Code is being remapped by a live experiment
Claude Code users noticed the tool reporting an effort value of 10 while set to high, and asked whether Anthropic had quietly cut effort. Thariq of the Claude Code team replied that the team sometimes tests API serving configs before rolling them out, and that one running now maps the numerical effort value differently. He said the scale is not 0-100, the number is not meaningful on its own, and the effort you selected is the effort you are getting, backed by in-depth evals showing no change in model performance. Sceptical users in the thread reported Opus 5 spending 43 minutes on a config-file update that previously took two.
The next model will no longer paper over your harness
Drew Breunig argues that Fable ended the free lunch in model upgrades, the way the end of Moore's Law ended it for CPU performance. Before Fable, he writes, it felt silly to spend much time improving a coding harness or context strategy, because a new model would land at the same price or cheaper and paper over most of the problems. Fable's cost broke that pattern: GLM 5.2 shipped the same week at roughly a ninth of the cost and is good enough for most code, so he now shapes a design in conversation with Fable before handing a brief to GLM. He adds that Fable's access controls, dynamic degradation and required data retention push organisations to reconsider where they send data.
Work that runs for minutes, not one round trip, is what MCP now builds for
An updated Model Context Protocol roadmap names five priority areas for the next specification and beyond. First is agentic messaging primitives, covering server-initiated events and webhooks so clients are not left polling for results, and maturing the Tasks extension so it can move into the specification proper. The rest are unifying on an HTTP-native transport, agent identity and enterprise security through Demonstrating Proof of Possession and Workload Identity Federation, standardised tool-result handling with progressive tool discovery, and SDK ergonomics and conformance. The Server Card Working Group continues on .well-known metadata conventions, so a server can be discovered and reasoned over without connecting to it.
Screenshots go into the same cheap agent endpoint you already call, at no extra cost
DeepSeek released deepseek-v4-flash-vision-exp on 21 August, an experimental vision build of V4-Flash. It takes images alongside text on the same API endpoint developers already use, with no price premium, and matches the base model on pure-text agent and reasoning work. It is a sparse mixture-of-experts model with 13B active parameters of 284B total, and an image costs at most 384 tokens. The vision variant has to be requested by name, and images are accepted in user messages only.
One prompt can now stack several llm templates, and embeddings take a per-call key
Simon Willison released llm 0.33 on 22 August. The -t/--template option can be repeated so several templates combine in sequence, and llm embed and llm embed-multi accept a --key flag that passes a per-call key straight to embedding plugins without touching shared model state. Reasoning-capable Responses API models gain a reasoning_summary option taking auto, concise or detailed. The release also moves to the OpenAI Python library 3.x and swaps the httpx dependency for httpx2.
eBPF, DWARF and custom crypto are inside reach of developers who never learned them
Armin Ronacher argues that agents have changed which technologies a developer can take on. Familiarising yourself with a language matters far less when an agent can rewrite code across languages, and he sees Rust and Zig gaining ground partly because you no longer need to know all the tricks yourself. He points at people doing serious work with DWARF files, eBPF, custom network drivers and custom crypto, work he says was previously off-limits for lots of developers. His examples include Cloudflare Artifacts running a pure-Zig Git-protocol engine compiled to a roughly 100 KB WebAssembly module, and Vercel's fx, a Zig coding agent advertised as small and fast.
