Engineering leadership
How do I prevent multiple AI agents from building the same feature?
Coordinate before execution and verify after it. Give every agent a bounded claim on one current work item, expose overlapping active claims, and compare new public code symbols with existing implementations.
Prevent overlap before code exists
A queue is not enough when several tickets describe the same intent. Keep tasks linked to a shared requirement or project, and make ownership and in-flight state visible to every agent.
Before an agent claims work, search active artifacts for the same outcome and inspect sibling work under the same parent. Reject or coordinate overlapping claims.
One bounded claim
Bind the agent to an explicit task, repository, scope, and current premise.
Shared visibility
Expose active owners, sibling tasks, and changed files before implementation starts.
Post-change comparison
Compare introduced public symbols with the existing codebase before merge.
Detect duplication at two layers
Intent duplication means two artifacts promise nearly the same outcome. Implementation duplication means two symbols express near-identical public contracts in different files. The repair differs, so report them separately.
Helen’s Foundry checks the artifact graph for overlapping work and the indexed codebase for near-identical signatures. Reviewers can merge scope, share an abstraction, or confirm that the similarity is intentional.
Do not over-centralize
Parallel work is useful when boundaries are explicit. The goal is not one agent at a time; it is one accountable claim for each piece of intent.