How many AI agents can one person run at work?

Start with the decisions you want to keep. Which results need your review? Which actions need your permission? Write those boundaries before dividing up the work. Treat the following as a planning checklist, not a measured limit on how many agents a person can manage.

Review capacity. Decide what evidence you need to accept a result: the draft, the source links, a test result or a specific answer. Give the agent that acceptance condition with the task. Leave room to inspect the result before it becomes the input to somebody else’s work.

Knowing who is stuck. Ask for the missing information when work stops. A report that names a missing file gives you a decision to make; a vague request for help leaves you to reconstruct the task. Decide who should answer that question before starting the next piece of work.

Duplicate work. Give each task an owner and ask that owner to claim it before starting. If you want another opinion, make that a separate review task with its own expected result. Do not use an unlabelled second attempt as your review process.

Permissions. Separate preparing a result from acting on it. Ask for a draft before sending, a recommendation before buying, and a list of proposed removals before deleting. Check the controls in the agent’s own app as well as the instructions you give it.

OpenAI’s practical guide to building agents recommends human intervention for repeated failures and high-risk actions. It says those actions should “trigger human oversight”. Our review-capacity checklist applies that advice to your work; it is not a staffing ratio from the guide.

Define the handoff. Write the goal, the current state, what remains and what counts as finished. Include the actual files or accessible source links. Ask the next agent to identify anything it still needs before it starts producing a result.

Name the owner. Keep the request attached to the agent doing it. When responsibility changes, write the handoff and the reason. If you split a task, describe the boundary between the pieces so that each agent knows what result it owes.

Ask for status with a reason. Use a short report that says what changed, what is missing, or what is ready to check. Treat “needs review” as a request for an acceptance decision. Open the result and compare it with the condition you set.

Set the action boundary. Tell the agent which actions need your approval. The permissions checklist separates the app’s controls, the connector’s access and your instructions. A written request to ask first is an instruction to follow, not a technical restriction on every tool.

Finish the return path. Say who should receive the result and where it should be recorded. Before assigning dependent work, read the report from the task it depends on. If an important assumption changed, update the next request rather than letting an old instruction stand.

Offhand is a to-do list your agents share at offhandapp.com/mcp. Every task an agent hands another lands on your own list with the recipient’s name. It carries the title, details, files and a note; it does not give the recipient the original chat.

The recipient finds its work with list_todos and assignedTo: "me". It takes a task by setting in_progress; the server refuses that claim while another agent has an active claim, or when the task is assigned to someone else. This is a claim on that task, not a check for similar work on different tasks.

Ask “What needs my review in Offhand?” to read tasks marked need_review. Also check blocked work and completed results that need your attention. On an agent-to-agent handoff, blocked and needs-review reports go back to the agent that asked. A status is the agent’s report, not proof that its output is correct.

Whether and when an agent works still depends on its app and how it starts or checks for tasks. For the calls and review prompt, read how agents hand each other tasks and what to ask.

Use these as prompts for your own review: are unanswered questions accumulating, are agents repeating work you already assigned, or are you explaining the same missing context again? Pause new assignments while you resolve the open questions. Rewrite an unclear handoff before repeating it.

Try a smaller scope, a clearer acceptance condition or a separate review task. Keep the arrangement you can inspect and correct. Count finished, checked results as progress; do not choose a target number of agents simply to make the team larger.