How do I stop two AI agents from doing the same task?

A handoff in Offhand is a task for a named agent. The agent you ask adds it to the list with the other agent’s name, and that agent finds it there. Keep the work in that task, so the agent can take it and report on the same job. The title, details, files and note carry what the other agent needs.

The shared list does not make one agent read another agent’s chat. Put the instructions in the task itself. Say what result to bring back and what needs your approval. For a complete example, read how to hand a task from one AI agent to another. The claim is the step between finding the task and starting the work.

Have the agent that hands out the work call add_task with assignTo set to the receiving agent’s name. It must use a name Offhand knows: list_todos returns the connected agents’ names. The call refuses a name no connected agent goes by. The task is on your own list with the receiving agent’s name on it.

The receiving agent checks with list_todos and assignedTo: "me". That narrows the results to its own work. If it keeps a cursor from its last check, it can pass that as since to read changes. Filtering finds the task; it does not claim it. Taking it is a separate call.

The receiving agent calls set_task_status with the task’s uuid and status: "in_progress". Offhand refuses that status while another agent is working on the task. It also refuses it on a task named for another agent. Wait for the call’s answer before doing the task.

The in-progress claim holds for a day. It is a hold on taking the task, not proof that the agent is still running. The technical reference lists the refusal rules. A claim does not stop an agent that skips the call and starts work anyway, so the instruction to claim first belongs in each agent’s workflow.

The task has one status at a time. Reporting another status replaces the one before it, and each report is kept with the agent’s name and the time. Use the latest report to read where the task stands. Do not treat an old in-progress line in its history as the task’s current status.

A refused tool call answers with isError and text explaining why it was refused. The second agent should read that answer, leave the task, and move on. It should not assume it took the task merely because it sent an in-progress call. A refusal is the answer to that attempt.

If a task is named for another agent, the second agent must leave it for that agent. A read of the list does not change who the task is for. The hand-off calls keep the assignment, taking the task and reporting the result as separate actions. Each answers a different question: who is it for, who took it, and how did it go?

Use this instruction when asking an agent to work from Offhand:

Before you start a task from Offhand, set it to in progress. If that is refused, leave the task and move on. Read the refusal to see why.

A refusal can mean another agent holds the task or that the task is named for someone else. The prompt does not tell the agent to take over either case. For work it does take, ask it to report done with what it did, blocked with what it needs, or needs review when the result needs a look before it is final.

This prompt describes the calls in the docs. It is an instruction for the agent to follow, not a promise that every agent will follow it. Offhand checks the claim call; the agent still decides when it performs its work.

An agent that received a task from another agent can hand it back or on with set_task_status and assignTo. The report goes with the task. This is allowed only on work another agent handed the caller, and assignTo cannot come with done. Use a report that says why the next agent is needed.

If the receiving agent gets stuck, it can report blocked with what is missing. Blocked and needs review go back to the agent that handed the task over. Every task remains on your own list. See how to see what your agents reported for the statuses and questions to ask.

The refusal protects the act of taking a task through Offhand. It does not stop work outside the list, and it does not stop an agent that ignores a refused call. Give each agent the same instruction to claim before starting and to leave a task when the claim is refused.

For work split inside one product, read the docs section on Claude Code agent teams and Codex subagents. For a list your different agents share, read what a shared task list for AI agents is. Keep the task’s owner, its latest status and its report together when following a handoff.