Good morning.

A new agent platform launches most weeks now, each one promising an assistant that lives in your phone or your Slack and runs the business while you sleep.

In May I told you which one I usually recommend, and after building a ten-agent team for a workshop in Austin last month, I'd give you different advice: keep the part of an agent you'd want to take with you, and rent the rest.

In today's issue: the agent team I'd run instead, the one folder every agent reads, and the 80/20 agent you can run in twenty minutes this morning.

—Sam

Let’s get into it.

Last month I ran a workshop in Austin for CEOs of companies doing $10 million to $100 million, and before I showed them a single agent, I put up one slide of the agent platforms everyone in the room had heard of:

Dots from OpenAI, Grok Bot from xAI, Muse from Meta, Instinct, Viktor, Poke, Lindy and Gumloop. New ones launch most weeks.

They're convenient, they run all the time, and they live where you already work.

In May, in (Why) You Need to Hire a Chief of Staff Agent, I wrote that Lindy is what I usually recommend for building one. Lindy is still a good product, and I'd still put plenty of jobs on a platform.

But at this point, there are a lot more options and they’re all basically the same. What matters is the structure and architecture underneath, which once you have it, you can replicate your agents on any platform of your choice.

The primary thing to keep in mind is this: what you gain in convenience, you lose in control.

Agent platforms are very convenient. But you won’t have full control. For some of you, that’s fine. For most of you, you want and need full control.

So, let me refresh your memory on the fundamentals of agents so we’re all on the same page:

An agent has seven parts (the next section names them), and most of them are yours: what it knows about your business, the instructions for the job, the limits on what it can touch, and the record of what it did. Those are worth keeping in files you own.

The rest, the connectors to your tools and the schedule that starts it, are fine to rent along with the model underneath, because you can swap any of them in an afternoon.

Claude Cowork and ChatGPT are rentals too, and that's fine, because a team written as plain files runs in anything that reads files.

When a better app ships next quarter, you move the folder and keep the team, along with every rule you set and every decision it learned from.

Before you put an agent on any platform, ask it four questions:

A platform that fails any of the four keeps the part of the agent that took you longest to build. Platforms come and go, too: You Don't Need 20 AI Tools covered one that announced it was closing.

The first agent in this issue takes about twenty minutes in Claude or ChatGPT: one file you write and one prompt. The full team is a download for Cortex members: ten agents that read the same folder, installed in about five minutes, with the rest filled in a few minutes at a time.

For a business of one, team means roles. Every agent is the same model reading a different job description, so you run the whole team yourself, and each agent drafts for your OK until a job has gone through ten times without a change and you've written it a rule.

There are two paths, and the choice comes down to which app you already pay for, because the kit has the same instructions twice: CLAUDE.md for Claude and AGENTS.md for ChatGPT and Codex.

Claude Cowork or Code (the Claude desktop app, on a paid plan) suits you if you live in Claude. You attach the kit folder to a project, connect Gmail, Google Calendar and your CRM under Customize, and put the weekly runs on scheduled tasks.

The ChatGPT desktop app with Work or Codex suits you if you live in ChatGPT. The ChatGPT app runs the kit in a Work project with the same folder attached, and Codex reads AGENTS.md on its own when you open the folder.

The quick decision rule:

The 80/20 below runs in a plain chat in either, with nothing installed.

Every agent you'll ever run has the same seven parts, whatever it's built in. May's How To Onboard An Agent Like A New Hire put most of them into one job description, and that still works for a single agent.

Role and Method together are the instructions in the own-or-rent table above, and the model sits underneath all seven.

Two of these matter more once you run several agents:

Method is the job written out as steps, so the agent does it the same way every time and you can see which step went wrong when it doesn't.

Record is the log entry every run leaves, which is how you check an agent without reading everything it read, and how another agent can check it for you, which is the whole job of the Keeper.

If you built the folder from The Cognitive File System for Maximum Production, you have most of this already.

What changes when ten agents read one folder is that each file answers one question, so an agent reads the two files its job needs instead of all of them, and two agents can't quote you different numbers because they found them in different places.

The kit calls the folder the Business Cortex, and its files split like this:

For a team, two of the files do more work than they did for one agent.

Only you edit rules.md, so no agent can widen its own access, even when it's sure it should.

And decisions.md stops one agent reversing something you decided in September, because before it recommends anything it checks whether you already settled it.

The 80/20 needs one of these files, priorities.md, so that's where to start.

The 80/20 Agent sorts everything on your plate into KEEP, HAND OFF, PARK or KILL against your 12-month goal, so no more than three things hold your attention this week.

It's the agent the kit runs first, because it needs one file and gives you something back the same morning.

Use a goal with one number and one date, three priorities for the quarter, and everything on your plate with rough hours.

This is Nadia's, from the example business in the kit (more on her below):

Leave an hour blank when you don't know it, and the prompt will label its guess. Keep the killed list even when it's short, because it's how the sort notices an old idea coming back under a new name.

Paste your file into Claude or ChatGPT, then this:

The order in step 4 is deliberate. KILL and PARK come first so the cheap cuts happen before anything competes for a KEEP slot, and HAND OFF comes before KEEP so a job that doesn't need you never takes one of your three.

Check the leak first. If the agent picked Traffic, the numbers it quotes should show leads below their normal range, and if they don't, nothing after that point can be trusted.

Then look at every hours figure, and fix any guess the agent presented as your own number. Last, read the questions it asked you, because that's where the real decisions are.

Change the list in your file: Keep, Hand off: Jess, and so on, and move each parked or killed item to the killed-or-parked list with today's date.

Next Monday the sort starts from that file, so a killed item that comes back gets caught, and a parked item returns when its condition comes true.

In Claude Cowork, save the file in a project and add a scheduled task for Mondays at 7:00 with the same prompt, starting it with Scheduled run: 80/20.

In ChatGPT, put the file in a project and set the prompt as a scheduled task for the same time. Either way, it reads the list you approved last week and shows you what moved since.

I ran the 80/20 Agent from the kit on Kestrel Lane Coaching, the example business that ships with it:

Nadia Kerr, a solo leadership coach for first-time managers at software companies, about $190k a year, with a VA eight hours a week.

Her list is the one in Step 1, and the agent also read the rest of her Business Cortex.

The run below is real: trimmed for length, with each cut marked, and nothing reworded. The kit's version of the agent calls the leak the constraint, so that's the word you'll see in it.

The leak comes first, with the numbers that prove it, and every keeper has a step with a day and a number with today's value.

It caught the podcast: Nadia killed Start my own podcast in September, and seven weeks later it was back on her list with a name. The sort parked it until she has two guests confirmed or decides it's a different idea.

It also stopped at three things it shouldn't decide on its own, the recording she'd already committed to, the sponsor reports she keeps for herself, and a website contract that cut against "not a priority this quarter," and asked her about each one.

The 6.5 hours a week came back without anyone new. Jess took the books because her role already covers QuickBooks, two agents took the recaps and the inquiry screening, and Instagram stopped.

Every agent reaches your tools through a gate, one level per tool, and every tool starts at read.

Moving a tool from read to draft is your call, once it's connected and you've checked it reads the right account.

Moving a job from draft to act within rules takes evidence: its last ten drafts went out unchanged, and you've written the rule in full, saying exactly what it may do, only when, and never what.

The kit's own example is Lead Enrichment filling blank fields on a lead record, only when the value has a cited source, and never overwriting a field that already has one.

Every raised gate goes in a log with the date and the reason, so six months from now you can see why an agent can do what it does.

Some things stay off at every gate: moving money, deleting anything, granting access, and contacting anyone on the list of work you keep for yourself.

How To Onboard An Agent Like A New Hire made the case for starting at read, and Let Claude (or ChatGPT) Run Your Email List ran the same gates on one job, your list.

The 80/20 runs the same way in every business.

What changes is what's on the plate and which agent pays back first once you've cleared it.

Run alone, every route ends with you. The agents draft, you approve in one sitting, and the Chief of Staff's morning page tells you which drafts are waiting, so what the team costs you is a short review each day.

With a team, people.md does the routing. A HAND OFF goes to the person whose role already covers the work, the way Nadia's bookkeeping went to Jess, and nobody gets a task their role doesn't cover.

At an agency, each account lead reviews the drafts for their own clients before anything reaches the founder, and the Strategist adds a seat for the team only when people.md lists more people than you.

The 80/20 is one agent out of ten. Cortex members get the rest of this issue, four more sections and the whole team as a download:

You should have a team like this running for you, reading one folder you own, so the sorting and the follow-up happen whether or not you remembered them.

You can run the 80/20 from this issue every Monday as it stands. If you'd rather have the whole team installed, Cortex subscribers get the Agent Team Kit as a download:

This is only available to paying Cortex subscribers.

Cortex is the premium monthly membership version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.

In addition to 3-4 full, unrestricted regular issues per month, you also get:

So, if you’d like to receive full issues and the monthly special Signals and Circuits issues, then you should sign up for Cortex, which is $50/month (and you have full archives of previous regular issues for as long as you’re a paying member).

One note on Signals and Circuits: You only get these special issues for the months you’re subscribed—which means if you skip a ~30 day period, about a month, you won’t get them for that month. There are no back-issues for past months, so now is your chance to lock-in these upcoming issues while you can.

After you’ve upgraded to Cortex, log in to the website and come back to this issue, and the Build Pack will be unlocked below.

Ready? Get the Build Pack, full regular issues, and the upcoming Signals and Circuits issues:

Cortex is a monthly premium version of Bionic Business, helping entrepreneurs grow and scale with AI and Agents. You can cancel at any time, but there are no refunds, for any reason. All sales are final.

One tap tells me where most of you run agents today. If you want to say more, the box after your vote comes straight to me.

Nadia's list had fifteen things on it, and two of them, Instagram and a podcast she'd already killed once, took three and a half hours of her week before anyone asked what either did for the goal.

Finding them took one read of the whole list against one number, which is the job the 80/20 does every Monday.

Write your priorities file before Monday, with the goal, the three priorities and everything on your plate.

Run the sort, and kill one thing before you add a single agent.

Talk soon,
Sam Woods
The Editor