Engineering blog
Notes from running agents like a team.
Opinionated, specific and written by the people building AgentCrew and using it daily. Subscribe by RSS or in the footer.
“Done” is a claim until there is evidence
Why an agent's completion message is the least reliable artefact in your pipeline, and how evidence-gated completion tied to commits and requirements fixes it.
One worktree per task: parallel agents without collisions
How isolating every agent task in its own Git worktree makes parallel coding agents safe, why integration should be an explicit merge, and why conflicts must block.
A prompt is not a specification
Coding agents build the wrong thing at full speed unless requirements come first. How stable IDs, acceptance criteria and revisions make plans agents can be held to.
Two approvals, not zero: humans in an agent pipeline
Full autonomy is a demo and babysitting is a chore. The right number of human interruptions in an engineering change is two: approve the plan, approve delivery.