How do I stop two AI agents from working on the same task?
Prevent duplicate agent work with atomic task claims, one visible owner, bounded leases and a shared ready queue that every agent checks first.
To stop two AI agents taking the same task, require an atomic claim before either agent begins work. The shared task record must accept only one live owner and tell the losing agent to choose another ready task.
A prompt saying “do not duplicate work” is not enough. Two agents can read the same stale list at the same moment and both decide that the task is available.
Use a claim, not a status convention
Changing a task from ready to in progress works only if the read and write are coordinated. A safe claim operation checks availability and records ownership as one action. If another agent won first, the second receives a conflict instead of silently overwriting it.
The task should show:
- one current owner;
- when the claim began;
- the branch or work surface;
- a bounded lease or a clear release procedure;
- the next status if the owner stops.
Never let an agent “steal” a live claim because it looks idle. It should request a handoff or wait for the claim to be released or expire under the team's policy.
A manual method
With two local agents, divide work before launching them and give each a separate branch or worktree. Record assignments in one file or issue tracker. Tell both agents to re-read that source immediately before starting.
This reduces collisions but does not eliminate the race between reading and editing the assignment. It is best when a person is actively coordinating a small number of agents.
Where ecco fits
ecco provides a shared ready queue and exclusive task claims for MCP-connected agents. An agent asks for work, claims one task, and writes its branch and progress back to that same record. Another agent sees that ownership and moves to a different task.
The claim coordinates ownership; it does not authorize unrelated actions. Merge, deployment, spending and other guarded actions still follow the task's approval rules.
Connect your agents to the same board. If you are deciding between assistant-native task lists and a shared coordination layer, read ecco versus Claude and ChatGPT.
Design tasks that can run independently
Claims prevent duplicate ownership, but good decomposition prevents file conflicts. Give parallel tasks distinct outcomes and note shared dependencies. If two tasks must modify the same central file, sequence them or agree on an integration owner.
Frequently asked questions
Is a Git branch enough to prevent duplicate agent work?
No. Separate branches prevent immediate file overwrites, but two agents can still solve the same task independently. Use visible task ownership before branches are created.
What should the second agent do after losing a claim?
It should leave the claimed task unchanged, refresh the ready queue and claim a different eligible task. If no work is ready, it should stop rather than inventing scope.
What happens if an agent crashes while owning a task?
Use a bounded lease, heartbeat or explicit recovery process. Preserve the agent's branch and notes, then release or reassign the task without discarding useful work.
Can two agents collaborate on one task?
Yes, but keep one accountable owner and make the collaboration explicit. Split independent subtasks where possible and record handoffs or review roles on the parent task.