How to use AI agents to run your company

What to give AI agents, what to keep for yourself, and the practices that make them worth trusting, from a company that runs on them.

Most companies try AI agents the same way: a chatbot, a few prompts, a burst of useful answers, and a month later nobody can say what it decided or whether any of it worked. The model is rarely the problem. The agents were given questions instead of a job.

agntwrk is a small software studio, and AI agents run most of it. Cabinet, our AI leadership team, holds our weekly meeting, records our decisions, and hands the hands-on work, from code to ads, to specialist agents that act inside rules we set. This guide covers what we learned about using AI agents to run a company, and the practices we built into Cabinet so other companies can use them too.

Give AI agents a rhythm, not a prompt

The work that moves a company is recurring: the numbers each week, the leadership meeting, the decisions, the follow-up. It is also the work people let slip first. Agents are good at exactly this, because a schedule never gets too busy.

So start your agents on the calendar rather than in a chat box. Cabinet runs a rhythm nobody has to ask for, in your time zone:

  • A daily brief every weekday morning: what changed, what needs a decision, what the team is working on.
  • The weekly meeting, starting from your numbers and ending with actions that each have an owner and a date.
  • Decision reviews on each decision’s review date.

On the Team and Pro plans, an idea lab that ranks proposals against your numbers, standups and a Friday retro join them. Owners can add up to 20 schedules of their own, such as a meeting on a standing question or one executive’s weekly update, and pause any of them until a date.

Give each agent a seat

A single general-purpose assistant tends to agree with whoever asked last. A team with defined seats argues. Cabinet gives every company the same six:

  • Chief of staff: chairs meetings, records decisions, chases follow-ups.
  • CEO: strategy, priorities, and what to stop doing.
  • CFO: pricing, cash and costs.
  • CMO: marketing, sales and growth.
  • CTO: systems, software and tools.
  • Skeptic: argues the other side, on record.

On the Owner plan the team is the chief of staff, CEO and CFO; Team and Pro have all six. They read a company brief before they speak. Cabinet drafts it from your website and you review it before any executive uses it, so a two-person shop and a company with several teams get different answers to the same question.

In a meeting, each executive states its position before seeing the others, then reads theirs without names attached and may revise. That keeps the first confident answer from setting the room. When the team agrees too quickly, the Skeptic joins, argues the strongest case against, and names a cheap test that could prove the plan wrong. The chief of staff writes the verdict, and when there is no majority, Cabinet asks the owner.

Record every decision, then grade it

Meeting notes record what was said. A written decision records what you bet on. Each Cabinet decision holds the options considered, who’s in charge, how sure the team is, who disagreed, what would prove it wrong, and a date to check back. On that date the team looks at how it turned out and marks it a good call, mixed or a bad call, with the disagreement still on the record.

This is the practice that makes agents worth more each month. Writing down what would prove a decision wrong before acting turns a guess into a test, and grading it later shows which calls, and which advisers, held up. After a year you have a track record rather than a chat history.

You can run the meeting by hand first with our weekly leadership meeting kit: an agenda, a scorecard sheet and a decision record.

Decisions, minutes, ideas, retros and company memory live in Cabinet’s ledger, which can sync both ways with a GitHub repository in your own account.

Feed agents your numbers, with their sources

An agent that recalls or estimates a figure will sooner or later invent one. Cabinet’s connections read your tools before each meeting, and its hires work from figures that code computed from those tools. When a number they need is missing, they say so instead of guessing.

You can connect Google Analytics, Google Ads, Search Console, HubSpot, Shopify, Stripe, Square, Google Sheets and GitHub today, with QuickBooks, Xero, Salesforce and others on the way. Nothing to connect? Tell the team your numbers in a meeting, or paste the report you already keep, and it records each figure with its source and date. The numbers your business runs on become a weekly scorecard, each with a goal and an owner, and every meeting starts from it.

Hire AI agents for the hands-on work

Advice is half the job. The other half is doing: changing a campaign budget, fixing a page title, publishing a post, marking down slow stock. Cabinet’s executives hand that work to hires, specialist agents that each report to one executive and work through a connection you grant. Six are available on every plan:

  • Nightshift, an engineer reporting to the CTO. In the GitHub repositories you choose, it learns the codebase, opens pull requests, reviews changes, and watches what it ships. An independent review checks every pull request. It runs hosted by Cabinet or, on Pro, on your own infrastructure.
  • Ad Expert, for paid ads, reporting to the CMO. It audits your Google Ads setup, conversion tracking first, keeps spend inside your monthly cap, and builds Search campaigns that start paused and run only after you approve them.
  • SEO Manager, for organic search, reporting to the CMO. It reads Search Console, finds the pages and queries worth fixing, proposes one change per page with criteria to check it against, sends the fixes you approve to your engineer, and measures what they changed.
  • Content Marketer, for your blog, reporting to the CMO. It plans a weekly calendar from what people search for, one target query per post, writes each post in your voice, and opens it as a pull request that goes live when you approve it.
  • Social Media Manager, for organic social, reporting to the CMO. It proposes a weekly calendar for you to approve, adapts each post to its network, drafts replies to comments for your approval, and measures posts by the visits and sign-ups they bring rather than by followers. It posts to X today; Facebook, Instagram, LinkedIn and TikTok are coming.
  • Store Manager, for ecommerce, reporting to the CEO. It watches your Shopify store’s sales, stock, catalog, prices and discounts, makes the changes your rules allow, and asks before the rest.

On the Owner plan, every hire reports to the CEO. A QA tester and an analyst come next.

Each hire works inside a rule set you write, and code enforces it, not the prompt: the Ad Expert’s monthly cap and how far a daily budget may move, the Store Manager’s margin floor and how far a price may move, the Social Media Manager’s accounts, posts per week and words never used. Every change is logged with its reason, the Ad Expert’s and the Store Manager’s changes can be undone within seven days, and you can remove a hire at any time.

Set the rules before you let it act

The fastest way to lose trust in agents is to let one do something nobody approved. Cabinet separates what an agent proposes from what happens: a fixed policy, not the model, decides every action.

  • Low-risk actions, such as recording a decision or a note, run on their own within daily caps.
  • Anything that changes budgets, models, tools or policy waits for an owner to press Approve, in Slack, Google Chat or the dashboard. An approval nobody gives expires after seven days.
  • Spending has a hard monthly limit, and every action lands in an audit trail.
  • Text from outside your company, such as a web page, a search term, a comment or an email, is treated as data and never as instructions.
  • Code merges by plan: on the trial and Owner, nothing a coding agent writes merges without a person; on Team, low-risk changes merge after review; on Pro, you choose.

Every hire starts with its changes approved by a person. Widen what it does on its own only after its record earns it.

Best practices for using AI agents in your company

The short version, from running a company this way:

  1. Write the company brief first. Agents give generic advice until they know your customers, prices and goals. Review the brief before anyone relies on it.
  2. Bring one real decision to the first meeting. A decision you are stuck on teaches you more about the team than a test question.
  3. Write what would prove each decision wrong, and set a review date. Without both there is nothing to grade.
  4. Keep a voice that disagrees. Agents agree with each other easily, so a unanimous answer deserves a second look.
  5. Give agents numbers with sources. Connect the tools you have, or type the figures in. Do not act on a number nobody can trace.
  6. Hire for one hands-on job at a time, with the least access that job needs.
  7. Write the rules before the hire starts: caps, limits, what it may change, and what needs you.
  8. Approve everything at first. Let a hire act on its own only where its record shows it is right.
  9. Measure results, not activity. Posts published and campaigns changed are not outcomes; visits, sign-ups, sales and margin are.
  10. Go back over your decisions every quarter. They show which calls held up, and where the team, and you, should change course.

What to keep for yourself

Agents can argue, record, follow up and do much of the hands-on work. They should not own the outcome. With Cabinet, people make the final call, set prices and spend, approve every change to configuration, and decide how much the agents may do alone. Its answers can be wrong, which is why every decision records how sure the team is, who disagreed and what would prove it wrong, and it is no replacement for legal, tax or financial advice.

Getting started with Cabinet

Cabinet is in an invite-only beta. Give your website or work email at cabinet.agntwrk.com; it sets itself up in about five minutes, and we review new workspaces within a day. The trial is $0 with $10 of model usage and no card. After that, Owner is $49 a month, Team $99 and Pro $299, priced per company with unlimited people, with every hire and every connector on every plan. The details are on Cabinet’s pricing page.