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 agent stops keeping a todo list on Opus 4.8, Sonnet 5 and newer.
Claude Code 2.1.233 removes the todo and task-tracking tools — TaskCreate, TaskGet, TaskUpdate, TaskList and TodoWrite — from Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer models, and an environment variable brings them back. The release closes an NTLM credential-leak vector, where Windows paths spelled with the NT device prefix bypassed UNC path validation. It also fixes auto mode on Windows repeatedly stopping for manual approval on ordinary redirected Bash commands, a 2.1.232 regression, and reverts that release's Bash permission changes for Cygwin-style symlinks and input redirections. An opt-in memory cgroup for Bash commands on Linux means a runaway build cannot stall the session.
Open weights that fit one consumer GPU now score 73 on Terminal-Bench.
Alibaba released the Qwen3.8-27B weights on 14 August under Apache 2.0, the open checkpoint promised alongside Qwen3.8-Max. It is a 27.78-billion-parameter dense multimodal model with a 262,144-token context, and against Qwen3.6-27B it lifts Terminal-Bench 2.1 from 63.4 to 73.0, DeepSWE 1.1 from 13.3 to 42.2, SWE-bench Pro from 53.5 to 61.7 and OSWorld-Verified from 63.9 to 84.3. The FP8 build needs roughly 14 to 16GB of VRAM and runs under Transformers, vLLM, SGLang and TokenSpeed. Weights are on Hugging Face and ModelScope, and the model takes text, images and video as input.
The text your product gets back from Claude carries a detectable pattern.
Anthropic described on 14 August how Claude's responses carry SynthID-Text, a Google DeepMind technique that biases the choice between equally viable words so a cryptographic key can detect the pattern. Code is largely untouched, because exact output leaves no choice to bias, and comments are where it can appear. There is no way to turn it off, and Anthropic says light editing will not remove it while a full rewrite will. A detection API is planned, and a detected watermark says only that Claude was involved, not who owns the result.
One scheduler runs your CLI agents in parallel and replays the run byte for byte.
Bernstein puts no model in the coordination loop, so parallel runs in per-task git worktrees replay byte-identically, and each run can be recorded with a cryptographic audit chain that verifies offline. It is an Apache-2.0 scheduler with adapters for Claude Code, Aider, the GitHub Copilot CLI, OpenHands, Ollama, Devin Terminal and more than forty others, plus a generic prompt wrapper for anything else. Each task takes its own git worktree, which keeps the main branch clean. It installs with uv or pipx and is solo-maintained and still in beta.
