The agent kit from Sam Woods’s DealCon workshop: twelve agents set up for your company, a guided install, and a monthly refresh that keeps them working. One folder you own and control.
Up and running in five minutes in Claude Cowork or ChatGPT Work, on a paid plan.
Create a project, attach the kit folder, and paste one line. Four quick questions later, your first agent is working. Then add more context whenever you have a few minutes: every answer makes the agents sharper, and they ask for what they need along the way.
Move the CEO-Agent-Kit folder somewhere permanent, such as Documents. Not Downloads, and never from inside the zip: your agents save their work in this folder.
Open the Claude desktop app, go to Cowork, and create a new project. Name it after your company.
Add the whole CEO-Agent-Kit folder to the project: the folder itself, not the zip and not a folder inside it. Allow Claude to edit files in it when asked. Leave the project instructions empty: everything the agents need is in the folder.
Four quick questions, one at a time, then your first agent runs: the 80/20 Agent sorts everything on your plate.
The first sentence points Claude at the kit; the second is any phrase from the table below.
Each part takes about four minutes, in any order: add my customers, add my numbers, add my team, add my market, add my deals, or continue my install to go in order. Agents also ask for what a job needs.
In the sidebar, open Customize, then Connectors. Connect email and calendar first, then your CRM and accounting. Then say add my tools: it walks you through setting every send and write action to Block, so the agents can look but not touch.
In the sidebar, open Scheduled, then New task. Paste the agent’s prompt from routines/schedule.md, pick the time, choose the kit folder, and run it once by hand. Your computer must be awake with the app open. Then say record my schedule.
Move the CEO-Agent-Kit folder somewhere permanent, such as Documents. Not Downloads, and never from inside the zip: your agents save their work in this folder.
Open the ChatGPT desktop app, switch to Work with the toggle above the message box, and create a new project. Name it after your company.
Add the whole CEO-Agent-Kit folder to the project: the folder itself, not the zip and not a folder inside it. Allow ChatGPT to edit files in it when asked. Leave the project instructions empty: everything the agents need is in the folder.
Four quick questions, one at a time, then your first agent runs: the 80/20 Agent sorts everything on your plate.
The first sentence points ChatGPT at the kit; the second is any phrase from the table below.
Each part takes about four minutes, in any order: add my customers, add my numbers, add my team, add my market, add my deals, or continue my install to go in order. Agents also ask for what a job needs.
Connect email, calendar and your other tools from the app’s connector or apps settings. Where a connector lets you turn off or require approval for its write actions, do that. Then say add my tools and it checks each one.
Create a scheduled task from inside the project and paste the agent’s prompt from routines/schedule.md. Run it once by hand first. Keep the computer on and the app running. Then say record my schedule.
Menus move. These steps follow how both apps work in October 2026. If something looks different, ask in the project: “Help me attach the CEO-Agent-Kit folder with edits allowed.”
You don’t install the agents or skills one by one, and nothing goes in the project instructions. They’re plain files in the folder. Each chat starts with the one line that points the app at the kit; after that, a phrase runs the agent you need. You can read and change every file.
| To | Say |
|---|---|
| Start any new chat | Read CLAUDE.md and follow it. (In ChatGPT: AGENTS.md) |
| Get started, or add more | start my install · continue my install · add my customers |
| Get the morning brief | brief me |
| Decide what to drop | run my 80/20 |
| Think through a big decision | help me decide whether… |
| Clear what’s waiting for you | go through my approvals |
| Build a new agent | build me an agent that… |
| Check your agents this week | check my agents |
| Run the monthly refresh | refresh yourself |
| Check your connections | check my setup |
Each agent is a role written down in plain files: what it owns, what it reads, which tools it can reach and how far, what it may never do, and what it records.
One page before your inbox: your day, what moved, and whether your week is going to your three priorities.
“brief me”Sorts everything on your plate into keep, hand off, park or kill against your 12-month goal. Three keepers at most.
“run my 80/20”Runs a big decision through the right mental models, seats an advisory board that argues, writes a one-page memo, and checks back later.
“help me decide whether…”What changed this week at your competitors and at the companies you might buy, and which change matters.
“what changed in my market”The revenue lever worth testing and the three biggest cost leaks, each priced in annual EBITDA.
“run my profit check”Researches, scores and routes every new lead, and writes sourced facts back to your CRM without overwriting anything.
“work my new leads”After each meeting: who promised what, the recap draft, tasks for named teammates, and a note in the CRM.
“follow up my meeting with…”Keeps brokers, lenders and key customers warm, with a real reason to reach out and a note in your voice.
“who should I reach out to”Checks whether a job should be an agent at all, then interviews you one question at a time and writes it, with its own tests.
“build me an agent that…”The agent that checks your agents: a weekly check, and the monthly refresh that tests every agent and proposes exact fixes.
“check my agents” · “refresh yourself”Scores a CIM or teaser against your buy box, flags what doesn’t tie, and writes the questions for the broker.
“screen this deal”Scores how ready you are for a buyer’s diligence, names what a buyer would discount, and turns the gaps into a 90-day plan.
“how ready am I to sell”business-cortex/: the files every agent reads first, filled in by the installlibrary/: what an agent is, how context works, writing agent instructions, when to hire helptests/: an invented test company and test cases, so agents can be run and graded without touching your datalog/, approvals/, outputs/: what every agent did, what waits for you, and its finished workFor DealCon attendees: a private conversation with Sam about putting agents to work in your company, or in one you’re buying. Something he normally charges for and rarely offers.
How to get in touchThe same model gives you generic output with nothing to read, and output that sounds like your company with the right files in front of it. The kit gives your agents three kinds of context. The full guide is library/how-context-works.md in the kit.
Short, checked facts you stand behind: what you sell, who buys, what your numbers mean, your priorities and your rules. The install writes it; you keep it current. Folder: business-cortex/.
The evidence: P&L exports, contracts and policies, meeting transcripts, CIMs, a CRM export. Dated files, plain formats. Never passwords or personal files. Folder: business-cortex/sources/.
Your inbox, calendar, CRM and accounting, read at the moment an agent runs, through connectors. Use a connector for what changes daily and a file for what changes monthly.
A connector is an MCP server. MCP (Model Context Protocol) is an open standard that lets an AI app work with a tool through a fixed menu of actions, such as “search messages” or “create draft.” The AI can only do what’s on the menu, and you control each item.
Every tool starts at read in your rules.md. A job earns more only after 10 drafts in a row you approved unchanged, under a rule you write. Instructions ask; connector settings enforce. Use both.
Tools change, data goes stale, and agents drift. Once a month, say refresh yourself (or put it on a schedule) and the Keeper runs a full check. It reads and proposes; nothing changes until you approve the exact words.
Each connected tool: does it work, is it the right account, how recent is its data, and are its write actions still blocked to match your rules?
Business files past their date, exports that stopped arriving, and agents that missed their schedule or ran on old numbers.
Every price, person and priority an agent relies on, checked against your files today, and last month’s corrections read for patterns.
Real runs, not guesses: each agent does its job on an invented test company, sealed off from your data and tools, and is graded pass or fail. A test that passed last month and fails now leads the report.
A web check of what changed in Claude, ChatGPT, MCP and your connectors, and a re-check of every platform fact the kit depends on.
Keep, fix or retire for each agent, with the exact before and after for every change, waiting for your OK.
go through my approvals clears what’s waiting for you.Agents look before they touch, and each job earns more on evidence, one at a time.
Anything that reaches a customer, a lead, a seller or the public waits for you, shown in full.
What each agent read, what it decided and why, and what it couldn’t find. A run that didn’t log didn’t happen.
The agents prepare analysis for you and your advisors. They don’t give financial, legal, tax or investment advice.
Paste either one into ChatGPT or Claude. The free plan works. They’re simple versions of two agents in the kit; the full agents read your company’s files and run on a schedule.
Fill in the brackets, then paste.
You're my 80/20 advisor. My single goal for the next 12 months is: [goal, with a number and a date]. Below is everything on my plate right now: projects, initiatives, ideas, AI experiments. For each item, judge how much it moves that goal and how much of my attention it takes. Before you sort, tell me in one line what's most likely holding the goal back: traffic, conversion, retention, capacity, cash, or decisions. Items that work on that score higher. Then give me four lists: - KEEP: at most 3. The items that move the goal most per hour of my attention. - HAND OFF: matters, but shouldn't be mine. Say who should own it. - PARK: worth doing later. Name the condition that brings each one back. - KILL: stop now, with a one-line reason each. For every KEEP, name the first step I can take this week and the number that will tell me it's working. Be blunt. If my goal is vague, ask me one question to sharpen it before you answer. My list: - [item] - [item] - [item]
For one decision you’re weighing.
I need to make a decision: [the decision, e.g. "whether to acquire a competitor at 4x EBITDA with a seller note"]. My goal is [the goal it serves]. Act as my decision partner. Don't decide for me. 1. Restate the decision and give me at least three options, including doing nothing and one I haven't considered. 2. Ask me the three questions only I can answer. Wait for my answers before going on. 3. Then run the decision through these lenses, one finding each: - inversion (how would this fail?) - base rates (how do decisions like this usually go? Say when you can't source it) - incentives (who wants what, and why now?) - second-order effects (and then what?) - opportunity cost (what else could this money and time do?) - margin of safety (what if my key assumption is wrong?) 4. Make the strongest honest case against the option I'm leaning toward. 5. Run a premortem: it's 18 months from now and this failed. Give me the three most likely reasons. 6. End with: your recommendation and how confident you are, the two facts that would change it, my walk-away conditions, and the first step with the number to watch.
A hands-on session that closed the AI track at DealCon’s CEO M&A Summit in Austin. The idea was simple: leave with agents working for your company, not a page of notes.
Most AI talks stop at strategy. This one showed five agents at work on a demo company buying a competitor, then gave everyone the same agents to install the same day, along with the pattern to build any other and the system that keeps them honest.
Daring Robot is an applied AI studio. Since 2021 we’ve built, launched and run our own agent-run businesses, and that’s where most of our time goes. For DealCon attendees, we’re making room to help in four ways.
A senior AI lead in your leadership room. What to build, what to buy, what to ignore, and a roadmap your team can run without us.
A focused look at where agents will pay back across marketing, sales and operations: what to build first, what to leave alone, and the number each agent should move.
Hands-on workshops where your people build and run agents for their own work, with the limits and checks that keep them safe.
Where AI belongs in the value-creation plan of a company you just bought, and what to leave alone until the integration settles.
A one-on-one about where agents would pay back in your company, or in one you’re buying. It’s something I normally charge for and rarely offer. For the people in that room in Austin, I’d like to make the time.
Talk to Sam Woods (sam@daringrobot.com)A line about your company and what you’re weighing is plenty.