The Co-Operating Model: Humans and Agents as One System
The last chapter showed that the operating model is the problem. This chapter names the model that fixes it. A Co-Operating Model is the design that makes humans and AI run as one system instead of two disconnected tracks, so the work stops falling through the gaps between them. You’ll learn what a Co-Operating Model is, how it differs from a traditional one, and a spectrum for placing any operation in your company: Assisted, Directed, Supervised, or Autonomous. By the end, you’ll be able to walk your own operations and name which mode each one runs in today, and which it could move to.
Jesse was in a meeting with Sofia, his VP of Operations, going through the org chart seat by seat. They had a marketing coordinator who was good at organization. She could manage processes, track content calendars, coordinate subcontractors. But she hadn’t been revenue-positive for months. Most of her week went to content production, editing videos and formatting posts, work that took her hours and still wasn’t shipping fast enough. The work was piling up behind her. Not because she wasn’t working hard. The role, as designed, required one person to do what should have been running across several tracks at once.
Sofia and Jesse looked at the chart and asked the real question. What if we stopped designing this seat for a human doing linear production work? What if we designed it instead for a human directing AI workstreams (work that runs on its own toward a goal without a person re-prompting between steps, covered in depth in the Design chapters)? The coordinator’s job wouldn’t be to make the content. It would be to direct the agents making it, hold the quality, and ship it.
That conversation changed the shape of the role. It also changed how we thought about every other seat. The problem was never the person. It was the operating model underneath her. The company had never been designed for a non-human worker. There was no seat for an agent, no defined handoffs, no accountability for work a machine produced. Every piece of AI output dissolved somewhere between the tool and the person who needed it.
Most companies have never asked which parts of a job require the person and which parts don’t. They grow roles by layering tasks onto people. The work was never designed, only assigned. So the role collapses when the person leaves, and AI has no defined relationship to it when it arrives. The rest of this book is the discipline of asking that question on purpose, seat by seat, and building the structure that answers it. We call that structure the Co-Operating Model.
The whole book in one line.
Everything in this book reduces to one equation.
Co-Intelligence + Co-Operation + Rhythm = Compound.
Read it in plain language. When you combine human and machine intelligence (Co-Intelligence), design how the two work together (Co-Operation), and run that design on a standard cadence of review (Rhythm), the results Compound: not a one-time jump, but a return that builds on itself quarter after quarter.
This chapter covers the first two terms and the spectrum they sit on. Rhythm reviews and adjusts the design on a repeating cycle, and Compound is the return that follows; the rest of the book teaches both. What you get here is what Co-Intelligence and Co-Operation look like in practice, and a way to place your own operations.
Co-Intelligence is two kinds of intelligence.
Co-Intelligence is human intelligence and machine intelligence working the same problem, and neither one is enough alone. Start with the half the AI conversation keeps skipping. The human.
A person who has spent years inside a business carries a world model, the accumulated sense of what matters, what a decision costs, and what a number means against everything else they know. That model is what tells you a churn risk is real before the data shows it, which client will tolerate a slipped deadline and which won’t, and what “good” looks like with your name on it. Julie spent a career watching leadership teams hand that judgment off to a process layer and lose the one thing only the person inside the work can supply: the relational read no system holds. You can write down a rule. You can’t write down the instinct that knows when to break it.
This is the work the last chapter pointed at when it said to elevate the work for your people. Humans aren’t displaced by a Co-Operating Model. They move up to the judgment.
| Humans bring the intelligence of | Agents bring the intelligence of |
|---|---|
| Judgment under ambiguity | Memory at scale |
| Context, history, relationships | Pattern recognition across your data |
| Strategy, taste, ethics | Tireless synthesis from broad sources |
| Standards — knowing what “good” looks like | Speed — producing at a pace no human sustains |
Now the other half. The machine.
First, a distinction, and it’s less about what the AI is than how you set it up to work. Use an AI as a chatbot and you’re in a back-and-forth: you ask, it answers, and nothing else happens until you ask again. You’re in the loop on every turn. Use the same AI as an agent and you hand it a goal, give it tools and access, and turn it loose: it works through a sequence of steps on its own, checks its own output, and keeps going toward the goal without you prompting each one.
The model underneath can be identical. Claude is a chatbot when you’re typing questions into a window, and an agent when it’s wired into your systems to run a job end to end. What separates the two isn’t the intelligence; it’s the harness around it: the goal it holds, the tools it can reach, the limits a person set, and whether it’s allowed to run the loop on its own. (The Source and Design chapters cover how that harness gets built.)
An agent brings what the left column can’t. It remembers at a scale no person matches. It finds patterns across more of your data than anyone could hold in their head. It synthesizes from broad sources without tiring, and it produces fast. Connect it to your systems and it can reason against more of your company’s knowledge than any single employee carries.
But the agent sits idle until something triggers it. It has no agenda, takes no initiative, and has no felt sense of what’s at stake. It doesn’t know what quarter you’re in or what your biggest customer is threatening. It knows none of that until a human transmits it. Hand an agent a task with no context and no definition of good, and it stays busy while working from the wrong picture.
That’s what the Co in Co-Intelligence is doing. The left column is the world model. The right column is what becomes possible once that world model reaches a machine that never tires. Co-Intelligence is the combination, and the combination is the only part that compounds.
Co-Operation is who owns what.
Co-Operation is how you split the work between the two so every piece has one clear owner. The rule is short. Humans own outcomes. Agents own tasks.
Ownership here is specific. It means who is accountable for the decision, the quality, and the next move when something goes wrong — not who’s cc’d or who’s aware. A human always holds that accountability for the outcome. A human can delegate the tasks underneath the outcome to an agent or a team of agents, and the agent owns those tasks and the outputs they produce. But the agent’s ownership is always derived from a human’s delegation. The outcome stays on the human’s name. Organizations hold people responsible for outcomes even when the process runs itself.
| Humans own | Agents own |
|---|---|
| Deciding what’s worth solving | Research, drafting, first-pass analysis |
| Directing the work and pricing the cost | Decision support and recommendations |
| Owning the outcome and the risk | Documentation, capture, monitoring |
| Setting the standard and holding it | Executing the routine, on call, on time |
The pattern in the columns is the model itself: judgment and direction on one side, execution and capture on the other. The human directs the agents instead of doing the work faster. That’s the shift that changes the math.
The balance cuts both ways. A company that lets the agent become the decision-maker gets AI slop with a professional’s name on it. A company that keeps the human inside the drafting and the documentation has wasted the model in the other direction, paying for expensive judgment to do work a machine could learn. The human sets the standard and owns the call. The agent carries the rule-work. Neither side is optional.
A traditional model grafts. A Co-Operating Model is designed.
Hold those two ideas next to a traditional operating model.
The last decade of business software sold a graft. You had a worker; now the worker has a tool. The work, the roles, and the accountabilities barely moved. You just expected more output from the same structure, and usually the only thing that changed was the number of places people had to enter data.
Julie watched this in a non-AI setting. At a global company where she ran HR, the business absorbed a large acquired division, and every function’s first instinct was the same: graft the new responsibilities onto existing roles. The org chart expanded. The operating model didn’t. Within a year and a half the company had to restructure anyway, this time by actually redesigning the work. The function that emerged ran leaner and more consistently, because each role finally had a defined scope and the capacity to execute it. AI is the same choice at a different layer.
In a Co-Operating Model, the human and the agent are both inside the loop, and the work itself is re-shaped around them: the agent takes the parts that follow rules, the person keeps the judgment, and the handoffs between them are designed in rather than left to chance. That redesign is the whole difference from a graft, where the work stays the same and only the tools change.
It also decides whether better AI ever helps you. A model is only as good as the context and structure around it. A graft company that handed everyone a chatbot built none of that, so a smarter model just makes individuals a little faster. A company that has mapped its knowledge and designed its workflows built the foundation a better model plugs into, so each improvement sharpens the whole operation, not only the people using it.
That same foundation keeps the company’s knowledge from walking out the door. When what one person knows lives only in their head (the vulnerability the preface opened with), a single departure can undo years of investment. The Source chapter builds the knowledge layer that holds it instead.
The spectrum runs from assisted to autonomous.
So far the split has looked all-or-nothing. In practice there’s range. What varies is where the human sits relative to the loop — the cycle of generation, review, and adjustment the work runs through. The human’s position moves across that loop. The accountability doesn’t: a human owns the outcome at every point.
Four modes. Each one is real, and each comes from how SuperWebPros actually runs.
Assisted — the human is in the loop, every step.
This is AI as an assistant for one-offs, the moments you want a quick hand and not a standing system. You’re driving the whole way. You ask Claude (an AI assistant, the same kind of tool as ChatGPT) for a quick post about something you just thought of, create a task by voice while you’re away from your desk, or ask where a project stands and what’s on your calendar today. You prompt, it answers, you decide what to do with the answer. Nothing happens without you. This is also the world from the last chapter: a few tools, no coordination, a person in the loop on every action. Most companies sit entirely here today.
Directed — the human is at the top of the loop, every cycle.
The human directs the agents and reviews every output before it ships. SuperWebPros’ marketing lead steers four agent workstreams toward one piece of content: one drafts copy, one generates visual options, one pulls and synthesizes research, one produces video. She doesn’t write the caption or cut the clip. She decides what’s worth publishing, sets the direction, and reads every output against her standard. Four streams run at once; she’s the one read on all of them.
Supervised — the human is at the edge of the loop.
The agents run the cycle inside guardrails the human designed, and the human handles exceptions and the final review. SuperWebPros’ quote workflow runs this way, and what looks like one “agent” is really a small team: a coordinating agent hands work to specialists, one researching past quotes for similar jobs, one drafting the options, one checking the draft against the criteria the design set. (Agents can manage other agents, the same way a team lead runs a crew.) Two guardrails keep the team honest: a spend cap, and a readiness gate that checks whether scope and timeline are actually present. When information is missing, the workflow auto-pauses and files a task asking the human, rather than guessing.
The human reviews every final quote before it goes out, owns the customer relationship, and makes the calls the agents can’t read off the data: when to discount, when to walk away. The agents run the loop; the judgment stays human.
Autonomous — the human is above the loop.
The work runs continuously on triggers and timers, and the human owns the design and monitors the results, stepping in only on exceptions. SuperWebPros’ billing operation runs this way. Following rules the team defined, it advances payment states, recalculates project totals as work changes, and posts a financial summary to the team channel on a set cadence. No one kicks it off each time. The human’s work moved up, not away: someone wrote those processes, set the thresholds that decide what counts as normal and what gets flagged, reads the summaries each cycle, and steps in when a number looks wrong or a payment fails. The accountability didn’t disappear. It moved up to the design.
Two things to hold about the spectrum.
It’s a map, not a ladder. The right stop depends on the work, the stakes, and how mature your knowledge and design are. Some work should stay Assisted forever, either because a person belongs in every step or because it’s too episodic and one-off to be worth building a system around. And not everything belongs at Autonomous: push work there before the design can hold the weight and good operations break, while automating work that rarely recurs is just expensive over-engineering. The skill is moving the right operations rightward, on purpose.
And across all four modes, the split was designed on purpose, not stumbled into.
What you can do now.
You don’t have the full instruments yet; those come in the Design chapters. But you can sketch the picture today, and the sketch is what makes the rest of the book concrete.
- List five operations that eat the most time in your business: quoting, content, onboarding, billing, support, scheduling, whatever yours are.
- Mark where the human sits in each one today. Doing every step yourself (Assisted)? Directing agents and reviewing each output (Directed)? Handling only the exceptions (Supervised)? Just monitoring the results (Autonomous)? Be honest: for most companies, nearly everything lands in Assisted right now. That’s the starting line, not a failing grade.
- Circle the one or two where the most human time goes into routine, repeatable work: the rule-following a machine could learn, not the judgment only a person can supply.
- For each circled one, name the next mode to its right and one thing that would have to be true to move it there: better-documented knowledge, a cleaner handoff, a guardrail you’d trust. You won’t have the full answer yet. Naming the gap is the point.
Keep the sketch. Signal, Source, and the Design chapters turn it into a plan.
The next chapter is the Framework: six stages, run one Sprint at a time.
Reflection Questions
- Walk three of your operations through the spectrum. If you’re like most companies, all three sit at Assisted today, a person in the loop on every step. Does that match what you see, or is anything already further to the right?
- Take the operation that burns the most human hours on routine, repeatable work. Which mode could it move to, and what’s missing today that keeps it where it is?
- Which operation would you never move to Autonomous, no matter how good the technology gets? What does that tell you about where human judgment is irreplaceable in your business?