Anthropic runs 10-agent Bugcrawl scan
- Anthropic is testing Bugcrawl inside Claude Code, a new repository-scanning feature that hunts for software bugs and proposes fixes across projects. - Screenshots and reports say Bugcrawl can launch 10 parallel agents at once, then return ranked findings while warning users it burns tokens fast. - Anthropic has been shifting Claude Code toward parallel-agent workflows across engineering products. (anthropic.com)
A code audit is a structured search for mistakes in software, from ordinary bugs to security flaws that can expose data or crash systems. Anthropic is testing a Claude Code feature called Bugcrawl to run that search across repositories. (testingcatalog.com) TestingCatalog reported on April 25 that Bugcrawl appears in Claude Code’s side navigation with a repository picker and a warning to start on a small repo because token use is high. The same report says the tool is meant to analyze source code, surface bugs, and suggest fixes. (testingcatalog.com) Separate reports from Daily.dev and ComputerBase said Bugcrawl is being built to run 10 agents in parallel, turning one scan into multiple simultaneous passes over the same codebase. Those reports describe the output as a prioritized list of issues for developers to review. (app.daily.dev) (computerbase.de) That setup differs from a single chatbot reading files one by one. Anthropic’s Claude Code product page says its engineers already work by “managing multiple agents in parallel,” which matches the Bugcrawl design described in recent reports. (anthropic.com) Anthropic has been laying out the same architecture in public engineering posts. In a June 2025 post, the company said its research system uses an orchestrator-worker pattern, with a lead agent delegating work to specialized subagents operating in parallel. (anthropic.com) The company has also been pushing Claude Code as a system that reads whole codebases, makes multi-file changes, runs tests, and works autonomously across a project. A bug-finding mode that fans out across a repository would extend that pitch from writing code to reviewing it. (anthropic.com) Anthropic’s own research this year pointed in the same direction. A January paper from Anthropic researchers and collaborators said an agent using property-based testing could identify bugs in large software projects across the Python ecosystem. (red.anthropic.com) There are still limits to what is public. Anthropic has not published an official Bugcrawl product announcement on its newsroom or engineering pages as of April 28, 2026, so the feature appears to be in testing rather than general release. (anthropic.com 1) (anthropic.com 2) For developers, the immediate tradeoff is visible in the interface itself: broader automated inspection in exchange for heavier token spend. If Bugcrawl ships widely, Claude Code would be moving from pair programmer to code-audit coordinator. (testingcatalog.com) (anthropic.com)
