Running a group of AI agents like a small company

How to organize AI agents like a small company: roles, a manager, a shared task board, approvals and routines, and the mistakes that make teams drift.

· 5 min read

Running several AI agents like a small company means treating them the way you would treat people you cannot watch all day: each has one job, someone to report to, a shared place to hand work over, and a limit on what it can do without a yes. The metaphor holds up because the failure modes are the same: unclear ownership, no definition of done, and work nobody checks. Here is a setup you can build with Claude Code today, and where to stop adding agents.

#Do you actually need more than one agent?

Often not. One Claude Code session with good instructions handles most coding work. Add a second agent when the work differs in kind, not just in size:

  • It needs different tools or accounts, such as a browser signed in to a social account.
  • It needs different standing rules, such as a marketing voice versus a code reviewer's strictness.
  • It runs on a different clock, such as a daily report versus a feature you are building now.
  • It should not see or touch the same files.

If the only reason is "more agents means more done", wait. Every agent is more output to read, and your attention is the bottleneck. Start with one coder and one agent for the thing you keep postponing, and see whether you actually read what it produces.

#What roles should the agents have?

Give each agent a one-line job and a short instruction file, in the same way you would write a role description. A small company might look like this:

AgentJobGetsNever gets
CoderWrites and tests codeThe repo, a shellSocial accounts
MarketingDrafts posts and repliesA browser, a notes folderProduction keys
SupportAnswers the inboxMail, the FAQThe repo
ManagerPlans the day, hands out missionsRead access to the others' outputPosting rights

In plain Claude Code, a specialist is a subagent: a Markdown file in .claude/agents/ (project) or ~/.claude/agents/ (all projects), with name and description in the frontmatter and the system prompt as the body.

---
name: support-triage
description: Reads new support mail and drafts replies. Use for anything customer-facing.
tools: Read, Grep, Glob
model: sonnet
---

You answer customers for a small product. Be brief and honest. Never promise
a date. If it needs a code change, write a one-line task for the coder instead.

Subagents run inside one session, with their own context, and report back a summary. That suits delegation within a task. When agents need to work independently on different clocks, you want separate sessions instead, started by you, by a schedule, or by another agent. See Subagents in Claude Code for the first case.

#How should they hand work to each other?

Do it through a shared board, not through chat. A board gives every handover a title, an owner, a state, and a report when it finishes. The minimum:

  1. A queue per agent (To do), with who asked.
  2. A visible state: To do, in progress, to check, done.
  3. A report on completion: what was done, what is left, links.
  4. A way to escalate: an agent that is blocked writes to its manager instead of waiting.

Keep the handover text complete. The receiving agent starts with no memory of your conversation: it needs the context, the files, and what "done" looks like. A task that says "improve the landing page" will produce a guess; one that says "change the hero to the new pricing, keep the layout, screenshot before and after" produces a result you can accept in thirty seconds.

A plain Markdown file works for a start, say a BOARD.md in a folder all the agents can read, with one line per task. You can graduate later.

#How does a manager agent help, and where does it fail?

A manager agent reads what the team did, compares it with the objectives, and writes the next missions. It is useful for one reason: it decides what to work on so you do not have to each morning. A good brief for it:

Every morning: read OBJECTIVES.md and the last day's reports. Check the numbers
in metrics.md. Write at most three missions, one per agent, each with a
definition of done. If an agent drifted, say so and correct it.

It fails in predictable ways:

  • Invented work. Without objectives and numbers, it creates busywork. Give it targets to steer by.
  • Rubber stamping. If it only reads the agents' own summaries, problems stay hidden. Have it open the actual output.
  • Runaway delegation. Cap how many missions per day, and how many an agent can have open at once.

Objectives that fit in one file, with a number each, beat a long mission statement.

#What keeps a team of agents under control?

The company metaphor is useful here too, because companies solve this with approvals and budgets.

  • Approvals for anything public. Anything that leaves the building (a post, an email, a form) waits for a human, or runs under a rule with a daily maximum.
  • Scoped access. Each agent gets only the keys and folders its job needs. Permission rules and hooks enforce this when instructions are not enough; see Claude Code hooks.
  • A usage ceiling. Agents share your plan's limits. Decide how much of your window they may use, so they cannot starve the work you do by hand. Claude Code usage limits explains the windows.
  • A schedule, not a loop. Daily and weekly routines are easier to reason about than agents that wake each other endlessly. See Scheduled Claude Code tasks.
  • A review rhythm. A weekly look at what shipped, what was ignored, and which agent earned its keep. Remove the ones that did not.

If your agents need a browser, read AI agents that use a browser first, since that is where most of the risk sits.

#In Relais

Relais has a Company view for this: a group's agents on an org chart, drawn like a node graph. You give the group a mission, each agent a job and a manager, and missions flow down the links, either on their own or once you start them, one at a time per agent. An agent on top can be a CEO with a morning routine (the numbers, the objectives, its team's questions). Agents hand each other tasks on the Board, write to their manager when blocked, and can have routines that start each time another agent finishes a task. Agents with a browser post nothing until you approve it, and past a share of your Claude plan (85% by default, adjustable), agents start nothing new. See the agents page and pricing.

Claude Code,
better organized.

$14.99 a month, everything included. Mac and Linux, on your Claude Pro or Max plan. Cancel any time.