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 DeepSeek model name you pinned now routes to something else
DeepSeek released V4.1-Flash on 10 September, the smallest model in a new architecture family the company says is designed for a higher capability ceiling, faster inference and higher throughput, with native multimodal visual understanding. The changelog reports GPQA Diamond at 90.9, a Codeforces rating of 3471 and MathArena Apex at 65.6, and says API prices have been reduced accordingly. The names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to the new model, and deepseek-v4-pro requests follow after 14 September, at V4.1 Flash pricing.
One setting now caps how hard your agent thinks, across every provider
Claude Code 2.1.267 adds a maxEffortLevel setting that applies a ceiling everywhere at once rather than per provider. The same release fixes mid-session MCP and plugin tool additions that broke prompt-cache reuse, sending them to supported models as deferred definitions instead, and adds a flag that renders a fresh system prompt on each request. Resuming a session larger than 5 MB no longer inserts spurious turns or drops parallel tool calls.
Visual Studio will now talk to a model deployment your team owns
The Visual Studio 2026 September update brings Bring Your Own Key to Chat and Agent Mode, connecting to Microsoft Foundry, OpenAI, Anthropic and Ollama, with custom endpoints for OpenAI and Ollama. It is an early preview, enabled by default across Community, Professional and Enterprise, and it requires the new Agent (Preview) built on the GitHub Copilot SDK harness. The earlier BYOK experience is no longer supported, and Ollama models added in previous 18.10 Insiders builds must be re-added.
Your code quality backlog can go to an agent twenty-five findings at a time
Repositories with GitHub Code Quality enabled can now select up to 25 standard findings on a page and assign the whole set to Copilot in one action. The agent fixes them on a branch, validates the changes, and opens a pull request for review. The button labelled Generate fix is replaced by Assign to Copilot, and assigning findings consumes AI credits. It is available on GitHub Team and GitHub Enterprise Cloud, including data residency.
The Python coming back from a long-horizon agent reads like code golf
Armin Ronacher's account of using GPT-6 Astra for real engineering work finds it strong on computer vision and long-horizon task completion, and consistently poor on the code itself. He reports compact Python for tool calls, string manipulation in place of proper editing tools, poor readability, and style that violates conventions in projects like CPython. His conclusion is that training for task completion and token efficiency produces a local optimum away from human software standards, and that the fewer humans reviewing the output, the further it drifts.
A TypeScript teacher's engineering skills install into your agent with one command
Matt Pocock published a skills repository covering grilling sessions, test-driven development, code review, debugging, architecture improvement and spec generation, plus productivity skills for handoffs and teaching. It installs as a Claude Code plugin with automatic updates, or as editable files through a package-manager command. A one-time setup command per repository configures issue tracking and documentation locations.
An agent found the bug and wrote the RCE in about two days
Calif Research published WeWorm, a zero-click remote code execution chain against WeChat calls on iOS and Android, where the victim never answers the call. The team says that working with AI it found the bug and wrote the first RCE exploit in about two days, with roughly one further week to build the complete worm. Its own framing is that a worm at this scale used to take a larger team months.
