The Fleet tests Docker's Coding Agent Sandboxes CLI, triages issues, posts release notes, and fixes bugs—all autonomously in CI.
The Fleet, a virtual team of seven AI agent roles built on Claude Code skills. Each role is a markdown file describing a persona and responsibilities. The agents test Docker's Coding Agent Sandboxes CLI, triage issues, post release notes, and fix bugs—all running in CI without human initiation.
Skills define what each agent knows; a Ralph-loop defines iteration. The cli-tester runs 52+ exploratory scenarios across 14 tiers on macOS, Linux, and Windows nightly. It loads a build-engineer skill to compile binaries and a project-manager skill to deduplicate findings before opening issues. The software-engineer role uses a worker/reviewer loop—Opus implements, Opus with 1M context reviews—and delivers PRs. All Fleet agents run inside Coding Agent Sandboxes.