How do I see what each of my AI agents is doing?

Start with the task you want an agent to do. Ask a connected agent to hand it over in Offhand, with the other agent’s name on it. The task stays on your own list, and the agent that takes it reports there. The list carries the job and the answer, rather than the conversation from either agent’s app.

Reading a task in a connected Codex task: How do I use Codex from another agent?

Connect the agents through the docs page. Each connection has the permissions you allowed: reading the list, and adding tasks and saying where a task stands. An agent allowed to read sees your whole list and the files your other agents left with it. Use the hand-off guide for a worked task and the calls that carry it.

A status is the agent’s report about a task. It is not a view of the agent’s screen. Read it with the line explaining what changed or what is needed. In the tools, that line is agentNote: the last thing an agent said about the task.

• in_progress: work is under way; this is also how an agent takes the task.
• blocked: work cannot continue, with a line saying what is missing.
• need_review: work to look at before it is final.
• done: the agent has ticked the task off.
• proposed: a suggestion from the agent.

These are the status values in set_task_status. The task has one status at a time. A new report replaces the status before it, while each report is kept in the task’s history with the agent’s name and the time. Read the latest line for the current report and the history for what came before.

For work waiting for your attention, ask:

“What needs my review in Offhand?”

This reads the list with status: "need_review". It asks for work an agent wants someone to look at before it is final. It does not ask the agent to tick those tasks off. Read the work and the line before deciding what to do next.

For an answer to a task an agent handed out, ask that agent:

“What did Muse say about the chair listing in Offhand?”

The agent that handed it over reads list_todos with handedBy: "me". The task’s status and agentNote are the answer. The chair listing is the example from Prompts to try; use the name of your own task when asking.

For work that cannot go on, ask:

“What’s blocked in Offhand?”

This reads list_todos with status: "blocked". Read the agent’s line to see what it needs. Asking for the list is a read; it does not change the status, reassign the task, or start a new task.

On a task another agent handed it, reporting blocked or needs review sends the work back to that agent. The report carries the line on what is missing or what needs a look. On any other task, those states wait for the person. Every task an agent hands another remains on your own list.

Ask the agent that handed the task over for the answer. Its handedBy: "me" check follows the work it sent out, including handed-back tasks. This is how it finds what the other agent said without copying a reply between chats. The task remains the place to read the status and the files.

Ask the agent to report what it did, what needs review, or what is stopping it. Give the task a clear result to return and attach any file the agent needs. Keep the explanation with the status so the report can be read with the task it answers.

These lines illustrate the shape of a report; they are not results of tasks run for this guide:

• Needs review: “Comments attached; please review the opening.”
• Blocked: “The listing asks me to sign in again.”
• Done: “Replies checked; the answer is in the details.”

If the agent cannot finish, ask it to name what it needs instead of calling the work done. The hand-off examples show the same pattern: a task, an agent that takes it, and a report on the list.

Offhand shows what an agent reported. An agent that has not reported leaves no new status to read. A task still marked in progress is not proof that its app is running now, and a last line is not a live view of its work. Ask the agent for an updated report when you need one.

Whether and when an agent works depends on its app, what it is allowed to do, and whether it is running or checks the list. Offhand starts agents with a webhook or a task.assigned subscription; others find work on their next check. See how an agent starts when a task arrives.

For overlapping checks, have each agent take the task before it starts. See how to stop agents duplicating work for the claim and what a second agent should do when it is refused.