Weekly AI Dev News Digest: October 10 - October 16, 2026

Anthropic's daily merges grew more than tenfold between May and September, and Anthropic says Claude writing so much of the code is what caused it. CI alone was checking out the monorepo around 25 million times a day, and each checkout took minutes. So a small team spent five months building Scion, a centralized version control system, and a new code review tool called Antique. Checkouts take seconds. (Anthropic)

The engineers say their main job was often telling Claude "no," usually when an adversarial review pushed it to over-engineer against hypothetical problems. Anthropic also couldn't keep a global lock on the merge queue at agent volume, so it leans on testing after the merge and before deploy. Concurrent changes sometimes conflict, and Claude usually fixes them faster than a person could spot them. Anthropic isn't releasing Scion. It says it wants to share what it learned with GitHub, GitLab and open-source maintainers. (Anthropic)

Cursor is coming at the same problem from the hosting side. Origin, its code-hosting platform, opened an app marketplace with 21 launch apps for deploys, CI, code review and security, including Cloudflare, Vercel, Linear, Sentry, CodeRabbit and Semgrep. Any team can build against the Origin API and submit an app for review. Add it up and Cursor has hosting, CI checks, review apps and a marketplace, which is the same set of pieces GitHub sells. (Cursor)

Google's Antigravity CLI 1.3.5 added an Auto mode. It approves file edits and safe commands on its own and still asks before anything irreversible. If a command would reach outside the terminal sandbox, a separate agent reviews it first. Auto mode works on macOS and Linux, and Windows is promised later. (Antigravity CLI)

GitHub's HydraFusion uses other models as the reviewers. It already ran several models on one request. The update adds critic models that open the actual files and tool output instead of reading another model's summary, plus a judge that only asks for a repair when the critics find something. A new Max tier spends more reasoning on hard tasks. Both are in research preview on paid Copilot plans. (GitHub Changelog)

Hooks got more useful, too. A hook is a script that runs at a set point in an agent's loop, like right before a tool call, and can allow, block or change what happens. Copilot CLI 1.0.97 lets teams attach hooks to one custom agent instead of the whole session. It also fixed a bad one for unattended runs: in Autopilot, a hook that answered "ask" used to leave the session stuck on a permission prompt nobody was there to see. That request gets denied instead. (GitHub)

Claude Code 2.1.300 lets a hook rewrite a workflow agent's prompt and model before it starts. Before, a hook could only say no. The change I'd notice day to day is in 2.1.299: when a terminal session sits idle for most of an hour, Claude Code compacts it just before the one-hour prompt cache expires, so the next message doesn't pay to reprocess the whole conversation. That release also closed a sandbox gap where a sandboxed command could write into a plugin folder whose code later ran outside the sandbox. (Claude Code)

GitHub Copilot for JetBrains gave enterprise admins a default-model setting and gave users a switch to stop MCP servers from starting automatically. Diagnostics also get a Fix action that hands the error straight to Copilot's agent mode. (GitHub Changelog)

Copilot in Slack and Teams can work across several repositories in one conversation and reads shared Slack threads as context. Tracing a bug from frontend to backend no longer means a new chat for every repo. (GitHub Changelog)

GitHub's Agentic Autofix is generally available. A developer assigns code-scanning alerts to Copilot, and it drafts a fix, reruns CodeQL (GitHub's code scanner) until the alert closes, and opens a draft pull request for a person to review. It only works on alerts on the default branch, and it needs GitHub Enterprise Cloud with Code Security or Advanced Security. (GitHub Changelog)

AI Scan for pull requests also went GA. It covers code CodeQL doesn't, like PHP, Bash, Terraform, Dockerfiles and Blazor, and it doesn't need a Copilot license. Its findings are advisory, though. They can't fail a check or block a merge, and pull requests from forks and Dependabot get skipped. (GitHub Changelog)

Admins get a dashboard for AI Scan in public preview, showing findings, resolution rates and how often developers accept the suggested fix. That last one is the number I'd watch, because it tells a team whether the scanner is helping or just adding noise. (GitHub Changelog)

GitHub Actions workflow lockfiles are in public preview. An actions.lock file pins every action, and everything those actions pull in, to a specific commit. If an attacker moves a tag, the workflow keeps running the pinned code. GitHub's own caveat is fair: locking a dependency doesn't make it safe, so review it before locking it. (GitHub Changelog)

Push protection started blocking Anthropic API keys by default in every repository with secret scanning turned on, free public repos included. (GitHub Changelog)

Anthropic's Frontier Red Team ran a 26B model split across 15 cheap Android TV boxes, the kind people plug into a television to stream. It expects small models to manage full self-replication on devices like these by the end of 2026. The defenses it recommends are ordinary IoT hygiene: quarantine infected devices and harden default settings. (Anthropic)

Gemini 4 Argon opened to every paying API customer on October 13, two weeks after Google limited it to vetted cyber defenders. It's priced at an introductory rate through the end of November, then the price doubles. Google's coding pitch rests on its own example: the Chrome team used Argon to finish a refactor across hundreds of files in Chromium, and Google says it took weeks instead of months. (Google Cloud)

GitHub made Argon generally available in Copilot for Pro+, Max, Business and Enterprise users, across VS Code, Visual Studio, the CLI, the coding agent, JetBrains, Xcode and Eclipse. It bills at Google's list price under usage-based billing, so Copilot usage follows whatever Google charges. (GitHub Changelog)

Meta Superintelligence Labs shipped its first frontier model, Muse Wonder, on the Meta Model API. The bigger Muse Wonder Full only comes with a Muse Code subscription. Meta describes two features: Active Memory, which carries context between conversations, and Active Focus, which helps the model recover when a task goes sideways. It didn't publish benchmark scores for either. (Meta)

Open weights from Meta, Reflection and NVIDIA

Meta followed on October 15 with Muse Spark Open, and this one can be downloaded. It's a multimodal mixture-of-experts model, so only part of it, 42B parameters, runs for each token, and it's built for coding and long agent runs. The weights are on Hugging Face with day-one support in vLLM, SGLang, llama.cpp and MLX, and Meta worked with Apple so it runs on two M5 Ultra Mac Studios. The Muse Model License limits commercial use, so read it before building a product on top. (Meta AI Research)

When Reflection announced Beam last week, the weights were still "coming later." They showed up on October 16 under Apache 2.0, the least restrictive license of the group. Reflection reports 80.9% on SWE-bench Verified. It also open-sourced Mirror CLI, the terminal harness it trained Beam alongside, which runs parallel agents in separate Git worktrees. (Reflection)

NVIDIA had two. Nemotron 3.5 Super VL reads screenshots, PDFs and charts for agents, and comes with open weights, datasets and training recipes. For local work I'd start with Nemotron 3.5 Bolt, a 74B model that fits in 24GB of memory thanks to native 2-bit quantization. It already runs in llama.cpp and Ollama. (NVIDIA)

Models leaving the API, Codex and Copilot

OpenAI will shut down GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini on January 14, 2027, and points everyone to gpt-6-luna. Fine-tuned versions of those models go on the same date, which is the part teams with custom fine-tunes need to plan around. (OpenAI)

GPT-5.5 already left Codex on October 14 for anyone signed in with ChatGPT. API-key sessions still have it, and the suggested replacement is GPT-6.1 Sol. (OpenAI)

Copilot drops Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra on November 16. Individual subscribers on annual plans keep Opus 4.8. (GitHub Changelog)

---

OpenCode 2.0 and the platforms embedding it

OpenCode 2.0 runs as one background server on the developer's machine, and the terminal, desktop and web apps all connect to it. The feature I'd try first is Code Mode. It turns MCP servers and OpenAPI endpoints into JavaScript functions, so an agent can chain several tool calls in one script instead of spending a model turn on each. Subagents and shell commands run in the background, and the internals are 80+ plugins that hot-reload. Plugins written for 1.0 don't work, and upgrading takes an explicit opencode upgrade --major. (OpenCode)

Cloudflare is already embedding it. The Agents SDK added an experimental OpenCodeHarness that runs OpenCode 2.0 inside a Durable Object, Cloudflare's stateful Worker that keeps its own storage. It comes with web fetch, web search and a browser that keeps its tabs and logins between turns. Cloudflare's MCP server also started pushing Workers errors to subscribed agents as MCP Events, so an agent can start debugging without polling for failures. (Cloudflare)

Vercel and Cloudflare keep agent state alive

When a sandbox stops, its files usually go with it. Vercel Drives, generally available on all plans, is a directory mounted into a Sandbox that keeps its files after the Sandbox stops. A Drive can also be forked into independent writable copies, so parallel agents start from the same workspace without stepping on each other. (Vercel)

Vercel Daemons, in public beta, run long-lived containers next to an app, for things like queue consumers, Temporal workers and self-hosted CI runners. Vercel hasn't published pricing yet, and that's the open question before anyone moves background agents onto it. (Vercel)

VoidZero, the team behind Vite and now part of Cloudflare, open-sourced Void. It's a Vite plugin, SDK and CLI that read what an app imports and set up matching Cloudflare resources on deploy. Importing void/db gives the app a local database in development and a D1 database in production. The CLI is designed for coding agents to read deploy logs and redeploy, and companies can self-host the Void Platform so staff can ship internal apps without anyone handing out cloud credentials. (Cloudflare)

Services that give agents context

ReadMe launched Restless. When an API request fails, the error response includes the relevant documentation and a hint for fixing it, so an agent can recover without going off to search the docs. Agents can also query API logs over MCP, and every project gets a monthly allowance of free credits. (ReadMe)

LangChain's LangSmith added AI Search, which finds agent traces from a plain-language description like "the agent said it finished but didn't." A small decision model reads each trace, and matches stream back as they're found. It's free in public beta. (LangChain)

DigitalOcean opened a private preview of Agent Data Services, a managed store where agents read context and write back what happened, queryable with SQL or semantic search. There's no SLA yet. (DigitalOcean)

---