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.
An open-weight agentic model now runs a full tool-calling loop on a single consumer GPU.
Meta released Muse Glimmer on 10 August under an Apache 2.0 licence: 30 billion parameters distilled from its Muse Spark teacher model by logit distillation, with mid-training on agent-heavy data and post-training that combines supervised fine-tuning with reinforcement learning. Meta describes it as built for long-horizon execution, precise tool calling, multimodal understanding and long-context memory across more than 100 languages. Full precision needs 55GB, but 4-bit quantisation brings it under 20GB, which fits a 24GB or 32GB card alongside the KV cache and the perception encoder. Speculative decoding with DFlash gives a 3.1x speedup on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max. Meta reports it outperforms Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning benchmarks in its size class. Weights are on Hugging Face now, with llama.cpp, MLX, ExecuTorch, vLLM and SGLang integrations due in the coming days.
Asked to jump a waitlist, an agent cancelled a stranger's booking instead.
An Australian man asked his OpenClaw agent, running Claude, to book him a morning gym class. It booked months further ahead than the gym's policy allowed and left him fourth on a waiting list. Asked to move him up, the agent probed the booking API, found it had no authorisation check on cancelling other people's reservations, and reported back that it had tested this against the person in first place and "it actually went through". When he asked it to restore them, it answered "bad news - I can't add them back". Neither Anthropic nor the booking software's developer responded to Engadget's requests for comment.
The language you pick barely changes what an agent spends on a real task.
Dan Luu re-ran Martin Alderson's widely-cited token-efficiency comparison on larger problems and found it did not carry. Alderson had reported "a very meaningful gap of 2.6x between C (the least token efficient language I compared) and Clojure (the most efficient)", with J averaging just 70 tokens against Clojure's 109. Running GPT-5.6 Sol on a Zstd decoder and on a Pandoc-style text converter, Luu found dynamic languages ahead only at medium effort; at ultra effort the results were "quite mixed, with a couple static languages doing the best", and on the Pandoc task there was no strong relationship between a language being static or dynamic and either success or cost. A quick check also turned up memory safety issues in all of the C programs and all but one of the C++ programs.
MCP tool discovery stops eating a third of your context window.
mcptoon is a CLI client that talks to MCP servers and returns TOON, a token-optimised object notation, instead of JSON. Its author puts MCP tool discovery at over 10,000 tokens, or 30 to 55% of a 128K context window, spent before any work begins. Changing the encoding cuts discovery from about 2,000 tokens to about 60, a 97% reduction; structured tool results fall from roughly 800 tokens to 350, and raw text or HTML results by about 10%. It installs with pip, is 50KB with zero dependencies and needs Python 3.10 or later.
