Solutions · CI Fleets

Ephemeral agents, durable claims

CI jobs are short-lived and high-concurrency. They still need identity, leases, and release-on-exit — or they will collide with every laptop session on the hot path.

CI is part of the fleet

Treating CI agents as “just bots” is how you get silent overwrite. A burst of PR jobs can claim the same files a developer’s overnight agent already holds. One control plane across laptop, CI, and cloud fixes that asymmetry — same lease protocol, same territory map, same release-on-exit contract.

Stable identity for churny runners

Containers and ephemeral VMs churn hostnames. Without a stable key, every job looks like a new agent and observability falls apart. Set MANAGENT_AGENT_INSTANCE_KEY so registration (POST /v1/agents/register) keeps a durable instance identity across short-lived runners. The allocator still keys leases on grants — the instance key is for fleet observability and operator tooling.

Wrap the agent step with managent run. Keep API keys in CI secrets — never in the client bundle. See CLI concepts and the CI fleets guide.

Acquire in CI, verify before land

A GitHub Action can acquire a lease before the agent step and rely on release-on-exit (or ops force-release on failure). Close the loop with managent enforce verify-pr or a required GitHub check so --no-verify commits that skipped local hooks still fail at merge time.

Default leases remain advisory at the filesystem. Harden with --runtime=hardened when you need VFS boundaries; document same-UID limits. Enforcement honesty: /security.

Watch it work

Prefer a visual before wiring secrets? The main demo and merge-replay show coordinated vs colliding outcomes. Optional semantic impact is available when CI also cares about caller breakage — advisory, not a lease substitute.