[{"data":1,"prerenderedAt":50},["ShallowReactive",2],{"blog-ai-agents-as-a-company":3,"blog-related-ai-agents-as-a-company":33},{"slug":4,"title":5,"description":6,"category":7,"categoryId":8,"keyword":9,"date":10,"minutes":11,"body":12,"html":13,"headings":14},"ai-agents-as-a-company","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.","Agents & automation","agents-automation","ai agents as a company","2026-11-17",5,"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.\n\n## Do you actually need more than one agent?\n\nOften 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:\n\n- It needs **different tools or accounts**, such as a browser signed in to a social account.\n- It needs **different standing rules**, such as a marketing voice versus a code reviewer's strictness.\n- It runs **on a different clock**, such as a daily report versus a feature you are building now.\n- It should **not see or touch** the same files.\n\nIf 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.\n\n## What roles should the agents have?\n\nGive 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:\n\n| Agent | Job | Gets | Never gets |\n| --- | --- | --- | --- |\n| Coder | Writes and tests code | The repo, a shell | Social accounts |\n| Marketing | Drafts posts and replies | A browser, a notes folder | Production keys |\n| Support | Answers the inbox | Mail, the FAQ | The repo |\n| Manager | Plans the day, hands out missions | Read access to the others' output | Posting rights |\n\nIn plain Claude Code, a specialist is a subagent: a Markdown file in `.claude\u002Fagents\u002F` (project) or `~\u002F.claude\u002Fagents\u002F` (all projects), with `name` and `description` in the frontmatter and the system prompt as the body.\n\n```md\n---\nname: support-triage\ndescription: Reads new support mail and drafts replies. Use for anything customer-facing.\ntools: Read, Grep, Glob\nmodel: sonnet\n---\n\nYou answer customers for a small product. Be brief and honest. Never promise\na date. If it needs a code change, write a one-line task for the coder instead.\n```\n\nSubagents 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](\u002Fblog\u002Fclaude-code-subagents) for the first case.\n\n## How should they hand work to each other?\n\nDo 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:\n\n1. **A queue per agent** (To do), with who asked.\n2. **A visible state**: To do, in progress, to check, done.\n3. **A report on completion**: what was done, what is left, links.\n4. **A way to escalate**: an agent that is blocked writes to its manager instead of waiting.\n\nKeep 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.\n\nA 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.\n\n## How does a manager agent help, and where does it fail?\n\nA 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:\n\n```text\nEvery morning: read OBJECTIVES.md and the last day's reports. Check the numbers\nin metrics.md. Write at most three missions, one per agent, each with a\ndefinition of done. If an agent drifted, say so and correct it.\n```\n\nIt fails in predictable ways:\n\n- **Invented work.** Without objectives and numbers, it creates busywork. Give it targets to steer by.\n- **Rubber stamping.** If it only reads the agents' own summaries, problems stay hidden. Have it open the actual output.\n- **Runaway delegation.** Cap how many missions per day, and how many an agent can have open at once.\n\nObjectives that fit in one file, with a number each, beat a long mission statement.\n\n## What keeps a team of agents under control?\n\nThe company metaphor is useful here too, because companies solve this with approvals and budgets.\n\n- **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.\n- **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](\u002Fblog\u002Fclaude-code-hooks).\n- **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.\n- **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](\u002Fblog\u002Fscheduled-claude-code-routines).\n- **A review rhythm.** A weekly look at what shipped, what was ignored, and which agent earned its keep. Remove the ones that did not.\n\nIf your agents need a browser, read [AI agents that use a browser](\u002Fblog\u002Fai-agents-with-a-browser) first, since that is where most of the risk sits.\n\n## In Relais\n\nRelais 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](\u002Fagents) and [pricing](\u002Fpricing).","\u003Cp>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.\u003C\u002Fp>\n\u003Ch2 id=\"do-you-actually-need-more-than-one-agent\">\u003Ca class=\"anchor\" href=\"#do-you-actually-need-more-than-one-agent\" aria-label=\"Link to this section\">#\u003C\u002Fa>Do you actually need more than one agent?\u003C\u002Fh2>\n\u003Cp>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:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>It needs \u003Cstrong>different tools or accounts\u003C\u002Fstrong>, such as a browser signed in to a social account.\u003C\u002Fli>\n\u003Cli>It needs \u003Cstrong>different standing rules\u003C\u002Fstrong>, such as a marketing voice versus a code reviewer&#39;s strictness.\u003C\u002Fli>\n\u003Cli>It runs \u003Cstrong>on a different clock\u003C\u002Fstrong>, such as a daily report versus a feature you are building now.\u003C\u002Fli>\n\u003Cli>It should \u003Cstrong>not see or touch\u003C\u002Fstrong> the same files.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>If the only reason is &quot;more agents means more done&quot;, 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.\u003C\u002Fp>\n\u003Ch2 id=\"what-roles-should-the-agents-have\">\u003Ca class=\"anchor\" href=\"#what-roles-should-the-agents-have\" aria-label=\"Link to this section\">#\u003C\u002Fa>What roles should the agents have?\u003C\u002Fh2>\n\u003Cp>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:\u003C\u002Fp>\n\u003Cdiv class=\"table-wrap\">\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>Agent\u003C\u002Fth>\u003Cth>Job\u003C\u002Fth>\u003Cth>Gets\u003C\u002Fth>\u003Cth>Never gets\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Coder\u003C\u002Ftd>\u003Ctd>Writes and tests code\u003C\u002Ftd>\u003Ctd>The repo, a shell\u003C\u002Ftd>\u003Ctd>Social accounts\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Marketing\u003C\u002Ftd>\u003Ctd>Drafts posts and replies\u003C\u002Ftd>\u003Ctd>A browser, a notes folder\u003C\u002Ftd>\u003Ctd>Production keys\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Support\u003C\u002Ftd>\u003Ctd>Answers the inbox\u003C\u002Ftd>\u003Ctd>Mail, the FAQ\u003C\u002Ftd>\u003Ctd>The repo\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Manager\u003C\u002Ftd>\u003Ctd>Plans the day, hands out missions\u003C\u002Ftd>\u003Ctd>Read access to the others&#39; output\u003C\u002Ftd>\u003Ctd>Posting rights\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>In plain Claude Code, a specialist is a subagent: a Markdown file in \u003Ccode>.claude\u002Fagents\u002F\u003C\u002Fcode> (project) or \u003Ccode>~\u002F.claude\u002Fagents\u002F\u003C\u002Fcode> (all projects), with \u003Ccode>name\u003C\u002Fcode> and \u003Ccode>description\u003C\u002Fcode> in the frontmatter and the system prompt as the body.\u003C\u002Fp>\n\u003Cpre class=\"code\" data-lang=\"md\">\u003Ccode>---\nname: support-triage\ndescription: Reads new support mail and drafts replies. Use for anything customer-facing.\ntools: Read, Grep, Glob\nmodel: sonnet\n---\n\nYou answer customers for a small product. Be brief and honest. Never promise\na date. If it needs a code change, write a one-line task for the coder instead.\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>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 \u003Ca href=\"\u002Fblog\u002Fclaude-code-subagents\">Subagents in Claude Code\u003C\u002Fa> for the first case.\u003C\u002Fp>\n\u003Ch2 id=\"how-should-they-hand-work-to-each-other\">\u003Ca class=\"anchor\" href=\"#how-should-they-hand-work-to-each-other\" aria-label=\"Link to this section\">#\u003C\u002Fa>How should they hand work to each other?\u003C\u002Fh2>\n\u003Cp>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:\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>A queue per agent\u003C\u002Fstrong> (To do), with who asked.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A visible state\u003C\u002Fstrong>: To do, in progress, to check, done.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A report on completion\u003C\u002Fstrong>: what was done, what is left, links.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A way to escalate\u003C\u002Fstrong>: an agent that is blocked writes to its manager instead of waiting.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>Keep the handover text complete. The receiving agent starts with no memory of your conversation: it needs the context, the files, and what &quot;done&quot; looks like. A task that says &quot;improve the landing page&quot; will produce a guess; one that says &quot;change the hero to the new pricing, keep the layout, screenshot before and after&quot; produces a result you can accept in thirty seconds.\u003C\u002Fp>\n\u003Cp>A plain Markdown file works for a start, say a \u003Ccode>BOARD.md\u003C\u002Fcode> in a folder all the agents can read, with one line per task. You can graduate later.\u003C\u002Fp>\n\u003Ch2 id=\"how-does-a-manager-agent-help-and-where-does-it-fail\">\u003Ca class=\"anchor\" href=\"#how-does-a-manager-agent-help-and-where-does-it-fail\" aria-label=\"Link to this section\">#\u003C\u002Fa>How does a manager agent help, and where does it fail?\u003C\u002Fh2>\n\u003Cp>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:\u003C\u002Fp>\n\u003Cpre class=\"code\" data-lang=\"text\">\u003Ccode>Every morning: read OBJECTIVES.md and the last day's reports. Check the numbers\nin metrics.md. Write at most three missions, one per agent, each with a\ndefinition of done. If an agent drifted, say so and correct it.\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>It fails in predictable ways:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Invented work.\u003C\u002Fstrong> Without objectives and numbers, it creates busywork. Give it targets to steer by.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Rubber stamping.\u003C\u002Fstrong> If it only reads the agents&#39; own summaries, problems stay hidden. Have it open the actual output.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Runaway delegation.\u003C\u002Fstrong> Cap how many missions per day, and how many an agent can have open at once.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Objectives that fit in one file, with a number each, beat a long mission statement.\u003C\u002Fp>\n\u003Ch2 id=\"what-keeps-a-team-of-agents-under-control\">\u003Ca class=\"anchor\" href=\"#what-keeps-a-team-of-agents-under-control\" aria-label=\"Link to this section\">#\u003C\u002Fa>What keeps a team of agents under control?\u003C\u002Fh2>\n\u003Cp>The company metaphor is useful here too, because companies solve this with approvals and budgets.\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Approvals for anything public.\u003C\u002Fstrong> Anything that leaves the building (a post, an email, a form) waits for a human, or runs under a rule with a daily maximum.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Scoped access.\u003C\u002Fstrong> Each agent gets only the keys and folders its job needs. Permission rules and hooks enforce this when instructions are not enough; see \u003Ca href=\"\u002Fblog\u002Fclaude-code-hooks\">Claude Code hooks\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A usage ceiling.\u003C\u002Fstrong> Agents share your plan&#39;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.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A schedule, not a loop.\u003C\u002Fstrong> Daily and weekly routines are easier to reason about than agents that wake each other endlessly. See \u003Ca href=\"\u002Fblog\u002Fscheduled-claude-code-routines\">Scheduled Claude Code tasks\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A review rhythm.\u003C\u002Fstrong> A weekly look at what shipped, what was ignored, and which agent earned its keep. Remove the ones that did not.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>If your agents need a browser, read \u003Ca href=\"\u002Fblog\u002Fai-agents-with-a-browser\">AI agents that use a browser\u003C\u002Fa> first, since that is where most of the risk sits.\u003C\u002Fp>\n\u003Ch2 id=\"in-relais\">\u003Ca class=\"anchor\" href=\"#in-relais\" aria-label=\"Link to this section\">#\u003C\u002Fa>In Relais\u003C\u002Fh2>\n\u003Cp>Relais has a Company view for this: a group&#39;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&#39;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 \u003Ca href=\"\u002Fagents\">the agents page\u003C\u002Fa> and \u003Ca href=\"\u002Fpricing\">pricing\u003C\u002Fa>.\u003C\u002Fp>\n",[15,18,21,24,27,30],{"id":16,"text":17},"do-you-actually-need-more-than-one-agent","Do you actually need more than one agent?",{"id":19,"text":20},"what-roles-should-the-agents-have","What roles should the agents have?",{"id":22,"text":23},"how-should-they-hand-work-to-each-other","How should they hand work to each other?",{"id":25,"text":26},"how-does-a-manager-agent-help-and-where-does-it-fail","How does a manager agent help, and where does it fail?",{"id":28,"text":29},"what-keeps-a-team-of-agents-under-control","What keeps a team of agents under control?",{"id":31,"text":32},"in-relais","In Relais",[34,40,45],{"slug":35,"title":36,"description":37,"category":7,"categoryId":8,"keyword":38,"date":10,"minutes":39},"ai-agents-with-a-browser","AI agents that use a browser: what works and what breaks","What AI agents with a browser do well (testing, reading, forms), where they fail (logins, CAPTCHAs, dialogs), and how to give one a browser safely.","ai agents with a browser",6,{"slug":41,"title":42,"description":43,"category":7,"categoryId":8,"keyword":44,"date":10,"minutes":11},"claude-code-headless-mode","Claude Code in headless mode: scripts, cron and CI","Run Claude Code without the interface using claude -p: pipe input, get JSON, pre-approve tools safely, resume sessions, and wire it into cron or CI.","claude code headless mode",{"slug":46,"title":47,"description":48,"category":7,"categoryId":8,"keyword":49,"date":10,"minutes":11},"scheduled-claude-code-routines","Scheduled Claude Code tasks: recurring routines on autopilot","Four ways to run Claude Code on a schedule: \u002Floop, cloud scheduled tasks (routines), Desktop tasks and plain cron. What each needs, and which to pick.","scheduled claude code tasks",1794931098788]