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.
Your subagents now start with the parent's conversation and prompt cache already loaded.
Claude Code 2.1.232 turns subagent forking on by default, and non-teammate agent spawns in interactive sessions now run in the background rather than blocking. Typing @ in the prompt mentions another live session by name, and SendMessage delivers to a bare name matching one live session instead of asking for a ref first. Sessions on one machine now keep unique names, gaining a suffixed variant when a name is already taken. The release also hardens the cross-session messaging socket directory on shared /tmp, refusing a pre-planted symlink or another user's directory.
Self-hosting closes on Fable 5 for coding, once the weights land in two weeks.
Zhipu released GLM-5.3 on 14 August, reporting a 50% gain in coding capability over GLM-5.2 from post-training scaling on an unchanged base model. It ranks first among open-source models on Terminal Bench 3.0 and Agents' Last Exam, with Terminal Bench 3.0 climbing to 28.3 from 4.6 and DeepSWE v1.1 reaching 66.9 from 46.2. On CyberGym it scored 84.5%, slightly ahead of Mythos 5 and GPT-5.6 Sol, while ExploitBench at 54.4% still trails closed frontier models on deeper exploitation work. Zhipu says the weights follow roughly two weeks after release, once safety evaluation is done.
Cloud agents boot from a prebuilt environment instead of assembling one each session.
Cursor's 13 August changelog introduces builds: pre-configured environment snapshots that Cloud Agents start from rather than setting up from scratch each session. Cursor reports internal environments booting ten times faster and three times faster time to first token. New environments get builds automatically, while existing ones need enabling from the Builds tab in the Cloud Agents dashboard, which shows status, logs and commit information. If a current environment fails, agents fall back to a previously successful build. Builds are included with Cloud Agents at no additional cost.
Every part of a new agent harness — model, sandbox, loop, UI — swaps out, source included.
DeepSeek put Harness into developer preview with source included under an MIT licence. The framework makes every capability a plugin: models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI can each be swapped or recomposed. Four runtime modes ship: Standard, with file editing, shell, file and web search, skills, planning, goals, subagents and workflows; Code mode, which exposes tools through an SDK so the model writes one TypeScript program combining multi-step operations; Minimal, a two-tool agent with persistent bash and a string-replace editor; and Creator, for authoring custom presets with runtime inspection. It runs from npx or a clone of the public repository.
A free coding agent installs from npm with an unreplicated 95.8% SWE-Bench claim.
Bullet launched on Hacker News as a coding agent built for latency, from a team that describes burning hours waiting on agent runs. Its site reports 95.8% on SWE-Bench Verified, a vendor figure with no independent replication. It routes tasks between fast and complex models by complexity, narrows code search rather than embedding whole repositories, and runs independent tool calls in parallel. Installation is a Mac download or an npm install of the CLI on Node 18 or later, currently a private beta on macOS and Linux, free with no key required to start.
