OctopusSign up

Resources

How to hire AI teammates

What they are, which jobs they can actually run, and how you stay in charge. Written like a founder talking to a founder.

Teammates

Jobs that survive

Give it the work that comes back as a draft

If the job recurs, and you still want a yes before it leaves the building, that is teammate work. Everything else can stay a prompt.

If it is not there on Tuesday, it is not a use case

A use case is not a capability. It is a job that recurs, starts from a trigger you can name, and leaves something behind you can check without a philosophy debate.

One marketing agent is four jobs that hate each other

SEO, social, email, and reporting do not share a rulebook. Averaging them is how you get a voice that is long, short, clever, and nobody at once. Split the lanes. Keep the brand file shared.

SEO is three jobs. Only one of them is writing.

One SEO agent is the wrong unit. Research, writing, and review do not share a brain. Review is the one that makes it safe, because a checklist is checkable and 'quality' is not.

Count the lanes before you hire a second AI

Tools are a shopping list. Lanes are jobs. The second teammate often adds work unless there is a Chief and a board. Start by asking whether the new job would force a context switch.

A billing reminder is a clock. Collections is a job.

A Zap will email on Tuesday. The real work is knowing whose billing day it is, preparing the invoice, nudging the right human, and escalating when nice stops working. That is Clerk, not a timer.

A prospect list is a sheet you can sort, not a chat

If the artifact is a paragraph of names in a thread, you do not have a list. You have a story about a list. Scout's job is criteria, research, enrich, then you approve outreach.

You do not want competitor alerts. You want a Monday brief.

Enterprise CI platforms, page-change watchers, and a teammate are three different products. Most founders actually want a one-page weekly brief on a short list of URLs, not a firehose.

How you run them

A screenshot is not proof of work

Clients do not want last week's status PDF. They want a live board they can open, without seeing your vault, your keys, or anyone else's account.

You do not babysit. You take the tap.

Human in the loop is not you watching the cursor. It is a column called Needs you, a mention that interrupts, and a send that does not leave without your yes.

The day they can touch Stripe is the day they can wreck it.

A teammate that cannot open your CRM, calendar, or inbox is still a chatbot. A teammate that can, with no scopes, is a leak with a first name.

If it cannot receive mail, it is still your chatbot.

A teammate without an address is a prompt box you carry around. Work cannot arrive. Follow-ups cannot wait. Replies cannot land with the one who started them.

Teach the procedure once. Stop pasting it every morning.

A skill is the job, written down, so Clerk's invoice format is still true when Maker needs the same number. You describe it like a hire. Chief writes it. You edit it like a document.

Chat memory dies. A workspace does not.

A thread forgets the vendor the second you close the tab. A shared desk of files, briefs, brand, and decisions is how Clerk and Maker still mean the same thing on Tuesday.

They need the live page, not last week's PDF.

Scout cannot watch a competitor from a screenshot you saved in March. Give the teammate a computer and a browser. Be honest about cost, rate limits, and what a page will actually give you.

If it only runs when you remember, you still have a chatbot.

Work on a clock. Weekdays at 9, Clerk runs invoices. Mondays at 7, the brief lands. Stagger the jobs so you do not stampede a limit. Send the output to you, not into a void.

Every model fills up or dies. Have a second one ready.

Providers hit a wall. Accounts reset. APIs take a nap. Ordered fallback is how Clerk finishes the 9:00 run instead of sitting in In flight until you notice.

You need a board, not another chat window.

People who try to run a team of agents always end up in the same place: a board, a feed of runs, a calendar, and a chat with the teammate. Watching logs is not the same job.

What it costs

Vs the session tools

Grok Bot is a teammate. You still need a desk.

xAI shipped named bots. The desire is a team. A team still needs a board, a yes before send, and a computer that is yours.

A chat with extra tabs is not a team. Here is the actual layer.

People started calling any pile of agents 'mission control.' Most of it is a session with better furniture. The ops layer is shared work that outlives the window.

ChatGPT Work is great at the artifact. It is not a Monday.

One agent, one doc, one sheet, one app: that product is sharp. A standing team is a backlog that is still there tomorrow. Use both. Do not make one pretend to be the other.

Cowork is a clean room. Then the room gets thrown away.

An isolated session is a gift when you want a sandbox. It is a problem when the job has to exist on Thursday. Lanes, a backlog, and costs should outlive the sandbox.

Codex is brilliant at parallel tasks. Parallel is not a team.

Disposable clouds, many jobs at once, code coming back. That is a factory for tasks. It is not Clerk, it is not an invoice lane, and it is not a backlog that still exists on Thursday.

Hermes is the closest native board. Then you have to live with the box.

Open-source kanban for agents, on one host, is the honest version of this category. The question is whether you wanted to babysit the box, or you wanted clients, email, and a fallback.

Claude Code is the deepest coding agent. Clerk should not live in a terminal.

Everything in Claude Code is a session, and the session is excellent. A persistent team is shared tasks, they wake each other, a live desk. Maker might use Claude Code. Mail should not.

OpenClaw is a strong personal runtime. A business still needs a desk.

Self-hosting a personal AI that actually runs is a real skill. Shared tasks, team chat, cost visibility, and a live board are a different skill. DIY versus hosted is the whole fork.

They all run one agent on one task. None of them keep the team.

Codex, Claude Code, Cowork, ChatGPT Work, OpenClaw, Hermes, Grok Bot: sharp at a sitting, or at a box, or at a bot on one account. The layer that stays alive between sittings is the whole product.