In 30 seconds
- Agents now write so much code that they break infrastructure built for humans: Anthropic replaced git with a system of its own.
- Anthropic released a Claude Code plugin that modernizes legacy code and delivers evidence that behavior did not change.
- Meta launched Muse Wonder, a model built for agents that competes with the leaders, although independent benchmarks place it just behind them.
[▲8] Anthropic replaced git with a system of its own: Scion
What it is: Anthropic went from about 2,500 merged changes per day in May to more than 30,000 in September, driven by code written by agents. Git and its mirrors could not keep up (a history of about 100 GB and around 25 million daily checkouts in CI), so a team of 7 people built Scion in five months and migrated the entire monorepo.
How it works: the server resolves history, blame and merges; the client only keeps a bounded cache. The result: checkouts about 100 times faster and infrastructure 200 times smaller than the git mirrors it replaced.
Why it matters: it is the first public blueprint of what development infrastructure looks like "at agent scale". Anthropic estimates that the rest of the industry will hit these limits within 6 to 12 months.
What to do: have the platform team read the article and start measuring the cost of CI checkouts (time and volume) as an early warning sign. Scion is not open source and is not available to third parties.
Source: Anthropic Engineering: Redesigning source control
[▲7] The code-modernization plugin: migrating legacy code with evidence
What it is: an official Anthropic plugin for Claude Code (v1.0.0, Apache 2.0 license) that analyzes a legacy system, extracts its business rules, proposes a plan, and only then migrates, either within the same stack (Java 8 to 17, Spring Boot 2 to 3) or across languages (COBOL to Java, PHP to Python).
What stands out: it does not just rewrite. It runs the old and the new code with the same inputs and compares the outputs, invents new edge cases, introduces a bug on purpose to confirm that the tests catch it, and delivers a verdict per module: PROVEN, PARTLY PROVEN or NOT PROVEN. The original code is never touched, and there is human approval at six key checkpoints.
Keep in mind: "proven" means verification evidence, not a formal proof. If the legacy system cannot run locally, the best possible outcome is PARTLY PROVEN.
What to do: try it this sprint on a small module. The analysis and planning stages change nothing; have someone who knows the business review the extracted rules before approving the migration.
Install: /plugin install code-modernization@claude-plugins-official
Source: plugin README
[▲7] Meta launched Muse Wonder, a model built for agents
What it is: two new models, Muse Wonder and Muse Wonder Full, with a 1-million-token context window and a focus on long-running agent tasks: memory across conversations and the persistence to retry or ask for clarification.
What the benchmarks say: Meta calls it "state of the art", but on the Artificial Analysis Intelligence Index Wonder scores 52 and Wonder Full 56, against 58 for Claude Opus 5.5. It competes, but it does not lead. The price going around (USD 4 per million input tokens and USD 20 per million output tokens) does not appear in the official announcement.
Availability: Wonder is on the Meta Model API and in Muse Code; Wonder Full is only for Muse Code subscribers.
What to do: if you use agents for long tasks, run the same real task with Wonder and with your current model, and measure whether it gets completed rather than going by impressions. The unofficial desktop app going around (muse-desktop) is two days old and has a single maintainer: it is not advisable for work.
Sources: Meta announcement · model card
[▲6] Meta's global AI hackathon: USD 1 million in prizes
What it is: a remote, worldwide hackathon from October 26 to November 4 to build with Meta's models. Each selected participant receives USD 150 in API credits. There are two tracks: independents (teams of up to 4) and startups.
Key dates: registration closes on October 25, or earlier if the cap of 20,000 selected participants fills up (selection is on a rolling basis). The official rules have not been published yet.
What to do: if someone on the team is interested, apply before October 20 with a concrete idea.
Source: event page
[▲5] WhatsApp Business MCP: managing WhatsApp from an agent
What it is: an official MCP server from Meta, in beta, that lets Claude Code, Claude Desktop, Codex or ChatGPT manage a WhatsApp Business account: phone numbers, templates, webhooks and test messages.
About the viral headline: it does not "kill chatbots". It is an administration tool for teams that already integrate the WhatsApp API, not an engine for customer service bots.
What to do: only if your company uses the WhatsApp Cloud API, try it in a development environment to set up templates and webhooks. Otherwise, you can ignore it.
Source: Meta documentation
[▲5] Repositories that help agents design better interfaces
What it is: a list of "10 repositories that turn Claude Code and Codex into designers" went around. On review, two links are broken and only three are tools for agents; the rest are component libraries.
The ones worth it:
- better-design: an MCP server with themes, design tokens and accessibility rules (WCAG).
- design-taste: 24 skills that teach agents aesthetic judgment.
- motion-primitives: well-maintained animation components, useful as a reference.
What to do: try the first two on an AI-generated interface to avoid the generic look these usually have.
AI Radar is a roundup of artificial intelligence and software development news, verified against primary sources. Ranking: estimated importance from 1 to 9; ▲ positive, ▼ negative, ● uncertain.