Keep this page open while you set up a standing team of AI agents. You act as the board: you set the goal, the limits and the approvals, and the agents do the work. Each item is one action you can check off.
A standing team differs from one chat session. The agents wake on a timer, remember nothing between runs and report through a task log. A tool such as Paperclip, an open-source orchestrator, can provide the parts below, but any tool with the same parts works.
0. Before you start
- Pick one job that repeats, for example a weekly report. Do not start with a whole company.
- Start with one agent. Add a second only after you have read the first agent's log.
- Check that the job fits a team (see "When not to use this" at the end). If it does not, stop here and use a single chat session.
1. Write a job description for each agent
Keep each description as a plain file you can open and edit. Tick all four parts for every agent:
- Role: write who the agent reports to and what it owns, in one or two sentences. One agent owns the plan; another owns visuals and nothing else.
- Startup checklist: write the steps it runs at every wake-up (see section 2).
- Persona: write how it thinks and speaks. This is also where you set voice and tone.
- Tool list: list exactly which tools and accounts it may use.
Treat the tool list as a set of keys. An agent can do only what its keys open, whatever its instructions say.
- Give each agent the least access it needs. A fact-checking agent reads drafts and cannot publish them.
- Choose a model per role. A cheaper model fits routine research; a stronger one fits the lead.
- Write a minimal version first. Edit a file when its agent misreads the job.
2. Set the heartbeat and plan for fresh context
The heartbeat is a timer that wakes the agent, lets it work and puts it back to sleep. Every wake-up starts with fresh context, so the agent does not remember its last run.
- Keep continuity in files and in the task log, never in the agent's memory.
- Write anything you want remembered into the instructions or into a task. A remark made in passing is lost.
Put this five-step startup checklist into every agent's instructions:
- Open my tasks that are still open.
- Read the comments left since my last wake-up.
- Check what changed since then.
- Check my own earlier work.
- Only then act.
- For a recurring job, set a schedule and an on/off switch. Read the run history after each run.
3. Keep one task log
The task log is shared memory for all agents and your audit trail.
- Use a tool with tickets. If you have none, create one file per task in a shared folder.
- Make every agent write its plan, its actions and its result into the task, not into a private chat.
- Steer by commenting on the task. The agent reads the comment at its next wake-up. Do not open a new chat.
- Read the log after the run, the way a board reads a report. Do not watch the agent live.
- Edit the instructions when the same mistake appears twice.
4. Set two brakes
- Budget cap: set a monthly spending limit per agent. Keep the limit outside the agent, so it cannot argue past it.
- Check spend per role: use the per-agent figures to see which role earns its cost.
- Approval gate: require your approval before any new agent exists. The lead agent sends a request with the role and the capabilities it needs; you approve or reject.
- Keep the gate on. Switch it off only after you have watched the team for a while and trust its judgment about roles. A team that hires freely grows faster than you can review.
5. Write a goal-level brief
Give the lead agent the goal, not the steps. Check that your brief has three parts:
- Goal: what you want.
- Expected outcome: which roles may be involved and what comes out at the end.
- Done condition: the point where the work must come back to you.
Example: "Nobody checks the texts before they go out. Hire an agent that checks the facts and the tone of every text before publication. Keep improving the results until you can come back to me and report that the copywriter and the new agent are both done."
- Leave out the steps: how to hire, which tests to run, in what order. The agent decides.
6. Install third-party skills safely
Skills and ready-made agent templates can come from public repositories with one pasted link. Your agents follow them with the keys you gave.
- Read the skill file yourself before you install it. A marketplace security badge does not replace reading.
- Check which access the skill needs. Refuse any skill that asks for more than the job requires.
- Attach it to one agent first, the one that needs it. Do not attach it to the whole team.
- Read the roles and tool lists of a "company" template before you import it. A team you did not design is a team you do not yet understand.
7. Run a weekly review
Set a fixed time each week and tick off these items:
- Read the task log for each agent: plan, commands, result.
- Check spend against each cap.
- Clear the inbox: approve, reject or answer every open request.
- Look for repeated mistakes and edit the instructions that caused them.
- Check that each tool list is still the minimum.
Ownership stays with you. A board does not do the work. It makes sure there is a goal, a limit, a record and someone who can say no.
When not to use this
Skip a standing team when:
- the job happens once, so there is nothing to repeat on a timer,
- one chat session can finish it, with you present the whole time,
- you cannot read the task log regularly, because an unread log is no oversight,
- you have not decided who answers when an agent gets it wrong.
In those cases, a single agent with a clear brief is easier to control.