The 7 Agent Rules I Use to Run My AI Team (And Why Executives Should Copy Them)

This article started as me talking into a recording.

No outline, no slides. Just my real world experience using AI agents. A little scattered, but talking it through let me create a flow of ideas, best practices, and real world stories quickly instead of worrying about grammar and structure. In about 15 minutes, I had 6 pages of notes.

Then I handed the transcript to my AI Chief of Staff, an agent that knows my background and how I like to write. It built a brief, pulled in outside research, and passed it to a second agent whose only job is drafting and editing. That agent put structure in place, checked grammar, and turned my flow into something useful for people to read. Then I went through the draft line by line in Track Changes mode and rewrote what didn't sound like me or needed more work.

My ideas and stories drove the process. The agents helped shape them.

I'm a few weeks into using multi-agent systems for my workflows, and I can see the real benefits and also the real risks. I've come up with a set of rules and best practices that I'm ready to share with any executive looking into these systems.

Why Agents Change the Game (and the Risk)

Earlier this year I wrote about how operations leaders can integrate AI in small steps that scale up to big results. Here we are in late 2026, and most AI pilots still haven't reached the bottom line. In McKinsey's 2026 State of AI survey, 37% of respondents report any EBIT impact from AI, essentially unchanged from last year, and only about 6% qualify as high performers.

One big reason is the gap between IT, who has to keep data locked down, and the people doing the work, who need access to it. Add confidentiality agreements, and nobody wants to be the one who exposed customer data to a third party. So pilots stall.

Here's what changed. Until recently, AI was mostly a chat: one input, one output. Now it comes as agents that work in the software world, in products like Grok Bot, OpenAI's ChatGPT Work, and Meta's Muse. Agents can open and change files, browse websites, log in, pull down data, and take actions on your behalf.

Those abilities are what finally make AI a useful tool and not just a fun chat. They're also exactly what makes agents risky. McKinsey's playbook for deploying agentic AI safely suggests thinking of agents as "digital insiders": entities that operate inside your systems with real privilege and authority.

Where the Agentic AI Trap Stands Now

In March I wrote The Agentic AI Trap, arguing that most executives should stick to discrete LLM tasks and simple automation instead of chasing multi-agent systems. That caution still stands. What's changed is that the tools have matured enough that a careful start is now worth it. The answer isn't a leap. It's guardrails, added one step at a time.

If you haven't already, you're going to hear more and more about AI agents. If you're refining your AI plans for 2027 and want to get ahead of the game, here's the playbook I'm running on my own work right now. I call it The 7 Agent Rules.

1. Match the Tool to the Job

Not everything needs an agent. For general questions and quick research, the old chat interface still works. Save agents for specific projects where you can give very clear instructions and set very clear parameters around your expectations.

Usage limits are real, and agents burn through them fast. Anthropic's engineering team reports that agents typically use about 4 times more tokens (the units these tools are metered in) than chat. Their guide to building agents puts the advice in one line: "We recommend finding the simplest solution possible, and only increasing complexity when needed."

On my team, every agent has a defined, repeating job. If a task doesn't repeat, it doesn't get an agent. The point of a multi-agent setup is to keep each workflow simple and well defined, which makes the whole system easier to manage.

2. Give Agents Their Own Sandbox

You have to give agents their own sandbox to play in. Before you connect anything, know where your agent actually runs:

• On your computer, fully offline: one of the safest options, because nothing leaves your device.
• On your computer, but online: give it one dedicated folder and access to only that folder.
• In the cloud: give it its own dedicated space for files, chats, and background, off your machine.

Where you can, give the agent its own accounts instead of yours. A separate account acts as a buffer.

My team works on its own cloud computer. Files never touch my PC unless I specifically download them. McKinsey recommends the same principle at enterprise scale: deploy agents "in self-contained environments with clearly defined network and data access."

3. Give the Least Access, and Clean the Data First

Least access is the underlying principle in all of it: limit access to only the things your agents are going to need to do their specific tasks.

Be very careful with connections to your email inbox, your airline login, or your bank account. If an agent gets a bad instruction while it has complete access, the risk is at its highest. The OWASP GenAI Security Project names the root causes plainly: "excessive functionality; excessive permissions; excessive autonomy." Its fix is to limit what an agent can call and do "to the minimum necessary."

A quick story I recently lived through. I use Amazon Web Services for some of my websites. A few months ago, in the middle of the night, I got a warning email saying I'd exceeded my daily budget and that my next bill was significantly higher. I checked my account, and it said I would owe about $30 million next month. You can imagine my panic running downstairs at 3 in the morning to find out what happened. It turned out to be a software glitch on their end, a bad update that sent everyone's estimated bills sky high. For those brief moments, though, it was real. Panic, sweating, increased heart rate, thoughts of bankruptcy running through my head. Carry that feeling with you when making decisions, because it only takes one bad software update, mistaken permission setting, or misunderstood command to generate a similar result.

Data needs the same care. Before an agent sees anything, strip out what's protected. For sales data, pull customer names, addresses, even product names. Keep quantities and prices, so the data is still useful but can't be tied back to anyone. Always consult a professional first. My list is a starting point, not an exhaustive one.

The safest start is no connections and no file access at all. The agent works only from public information and what you've given it. When it needs a third party, you make the contact and hand the answer back. My research agent works this way: public sources only, never logged in to anything. If it needs more, I go get it, and it picks up from there.

4. One Job per Agent, One Lead Agent

Set up one main agent, like a Chief of Staff, that understands all the background and acts as the gatekeeper. It farms out individual tasks to helper agents that each do one simple job.

My team is a Chief of Staff plus three helpers: research, editing, and repurposing. The editing agent is new every week and deleted after the article publishes, so it never drags a long history behind it. Any useful lessons it learns about how to edit better go into a permanent file that the next editing agent learns from.

The real benefit is how the work is organized. The complex, long-lived work (my background, strategy, and voice) lives in one place, with the lead agent and the files it reads. Simple jobs go to dedicated agents I can delete or duplicate often, so no simple job grows into a huge memory, and I never end up juggling several giant chat windows. Anthropic's engineering team describes the same pattern: "specialized sub-agents can handle focused tasks" with a fresh start, while "the main agent coordinates with a high-level plan."

5. Make Files the Memory

Every time you prompt an agent, it typically rereads the entire conversation first. The longer the conversation, the more work it does just to catch up.

The fix is a simple file system that acts like a memory for all of your agents: your background, the purpose of the project, the data, even "this is the way I like to communicate." Write it once, and you stop repeating yourself.

My setup has a voice file, a background file, a strategy file, and a START-HERE file that every new agent reads first, like onboarding for a new hire. Anthropic's team calls the idea "structured note-taking": the agent keeps notes outside the conversation and pulls them back when needed, because its working memory is "a finite resource with diminishing marginal returns."

One key tip: tell your agents to stamp every change with the date, what changed, and why, so you have traceability if you need to look into it later.

6. Keep a Human on Every Send Button

Anytime you're creating content or communication, it should be the human still driving it. The agents can help with editing and pulling in information.

A lot of people are using agents right now to pump out a ton of content and messages. Mass-produced content is a bad idea. People really only care about the stories, purpose, and unique perspectives that humans attach to a message. Get specific in anything you write, and focus on why you're taking the time to make it. Use agents to help you put structure around it, add data to it, and prepare it for distribution.

One of my best tips: after any customer meeting, the first thing I do is open an email and use voice transcription to speak all my thoughts into it. What happened, what we talked about, important points or actions, and any other idea that pops into my head. I email it to myself, strip out any protected information, and have an AI structure it and prepare it to share with others as needed. It always surprises me how much gets forgotten between leaving a meeting and getting back to your desk to write up notes. Don't wait. Get it out while it's fresh and turn it into an action plan fast.

Singapore's IMDA, in its Model AI Governance Framework for Agentic AI, lists "sending communications" and "making payments" among the irreversible actions that should hit a human approval checkpoint. Consumer agents are being built the same way: Meta says Muse "checks with the person before sensitive actions like sending an email or making a purchase." IMDA also warns about automation bias: rubber-stamping whatever the agent hands you. Keep approvals short and clear, and actually read them. Remember that you own anything that goes out, so take the time to do it right.

7. Leave a Paper Trail, and Start Small

As I mentioned earlier, whenever an agent changes something, put a timestamp on it. Why did it happen and why was it important? Timestamps carry the traceability idea from my March article over to agents. McKinsey puts it directly: agentic systems "should be created with traceability mechanisms in place from the outset." On my team, every file change gets a dated line: what changed, and why.

Then start small. Give the agent a goal and let it help you define how to get there. Stay high level at first, let an agent help design the system, and run small experiments. My rule for growth: add a new agent only after an existing one proves itself.

Your Agent Job Card

Before you turn on any agent, fill in this card. If you can't answer a line, the agent isn't ready yet.

Where Agents Help, and What Not to Hand Them

The way I think about it, agents can help your people work better, more efficiently, faster. Offload the repetitive work: recurring reports and data entry.

• Sales: analyze sales history and customers, then use public data to find new segments.
• Accounting: past due customers and other financial analysis.
• Marketing: monitor social and your website, track voice of the customer, and help draft and edit articles.
• Operations: trends, planning, quality defects and root cause, and ordering.

A simple operations example: tell your lead agent that every Monday you want to know what needs to be built this week, what got built last week, and what you've built historically for this month and the upcoming quarter. Say you have three helper agents, each analyzing its own data set. The lead agent pulls it together into one report Monday morning.

What I wouldn't hand them:

1. Personal communication and follow-up.
2. The final say on anything under your name.
3. Any irreversible action without your approval.
4. Anything physical. Agents are only digital for now.

Closing Thought: Start Small This Week

Pick one repetitive report your team builds by hand. This week, fill in the Agent Job Card for it, using only scrubbed data and no connections. You'll know it's working when the report shows up and nobody had to chase data for it. Check the report for the first few weeks to make sure it's running correctly before you let it run on autopilot, and go back to the change log for any errors or omissions.

I'm running this experiment on my own work right now, and I'll share what works and what breaks as it goes. If you want those updates, the newsletter is the best place to get them. Sign up below.