The Three Node Org Chart
The org chart of an AI company has three kinds of node.
Humans, agents, and hybrids. Four reorganising patterns that follow once you draw all three.
3
Node types
4
Patterns
8 min
Org audit
Most companies still draw only humans. The work doesn't compound.
A traditional org chart has one kind of node: humans in boxes. An AI-native org chart has three: humans, agents, and hybrids. Hybrids are humans whose primary work is to manage agents. Most companies still draw only the first kind. They have agents, but they treat them as tools owned by individuals. That's why their stacks are messy and their AI work doesn't compound.
Most companies running AI still have a 2018 org chart with AI tools sprinkled on top. Same boxes. Same reporting lines. Same job titles. Just with "powered by AI" added to a bio. That isn't an AI company. That's a company that uses AI.
The structure looks different when you rebuild around it. Three kinds of role. Two rules. Four patterns that follow from naming them honestly.
The shift: stop drawing your org with one kind of node. Start drawing it with three. Then apply two rules to every box.
The two rules
- Every agent has a human manager. No exceptions. Software without an owner produces orphaned outputs.
- No human does an agent's job. If a person is doing high-repetition, well-scoped, auditable work, you've miscast a role. Promote them, hire an agent, or both.
These two rules sound simple. They reorganise everything when applied seriously. The four patterns below are what happens when you do.
PATTERN 01 · HUMAN ROLES
Fewer humans. Each one more senior.
The next hire isn't a junior version of someone you already have.
When agents take over the repetition, what's left for humans is judgement work. That means fewer humans, but each one operates at a higher level. Junior copy-paste roles disappear. Senior decision-making roles expand.
This isn't a layoff story. It's a who you hire story. The next person you hire shouldn't be a junior version of someone you have. It should be someone more senior, paired with agents that handle what a junior used to do.
The trap: hiring a senior without agent fluency. The senior arrives, looks at the agents already running, and tries to do their work manually because that's the version of the role they know. The seat needs both: seniority and agent literacy. Filter for both, not just one.
Example: AUDIT THE NE X T ROLE YOU'RE ABOUT TO HIRE
You tell Claude: "I'm about to post a JD for a junior content coordinator. Walk it task by task. Which tasks are human judgement? Which are agent-shaped? Recommend a senior pairing if most of the JD is agent work."
WHAT THIS PATTERN PRODUCES
- Judgement work concentrates upward. Repetition concentrates into agents.
- One senior plus agents typically outperforms two juniors at lower total cost.
- The hiring cycle gets shorter. Fewer seats, but each one matters more.
WHERE IT BREAKS
- A senior without agent fluency under-delivers. Filter for both.
- Some judgement work can't be cleanly separated from the repetition it sits inside.
- "Promote everyone" reads good until the workload still needs juniors temporarily.
Use this when: you're about to post a new role. Audit the JD before it goes up. If most of it is agent work, the right answer is a different role.
PATTERN 02 · AGENT ROLES
Agents get job descriptions, too.
The chart shows them by name.
Agents have job descriptions. Agents have inputs and outputs. Agents have failure modes and escalation paths. In a serious AI company, that information is written down somewhere a human can find. Not lost in a prompt that lives in one person's head.
The bar: a new hire should be able to read the agent's chart the same way they read a colleague's. Name. Manager. Inputs. Outputs. What it can't do. What it escalates. When it retires.
The org chart should show agents the way it shows humans: with names, jobs, and managers. We have agent boxes in ours, labelled. The chart looks weird the first time. The work gets cleaner the first week.
Example: WRITE THE SPEC FOR AN AGENT ALREADY RUNNING
You tell Claude: "Take the agent that drafts our weekly newsletter. Read the last 6 issues. Write the agent's job description: inputs, outputs, failure modes, and escalation rules."
WHAT THIS PATTERN PRODUCES
- The agent stops living in one person's prompt history. It lives on the chart.
- Failure modes become first-class. Observable, owned, fixable.
- A new hire can take over the agent without onboarding from the previous human.
WHERE IT BREAKS
- A written spec ages. Weekly-touched specs stay alive. Quarterly ones drift.
- Most real failure modes only surface after 4 to 6 weeks of running.
- Specs don't replace observation. Someone still has to read the outputs.
Use this when: an agent has been running long enough to have failure patterns. That's the moment a spec earns its keep.
PATTERN 03 · HYBRID ROLES
The most leveraged seat didn't exist three years ago.
A human whose primary work is briefing and reviewing agents.
The hybrid is a human whose primary work product is briefs for agents and reviews of agent outputs. In our company, the operations lead spends roughly 60% of their time managing agents: writing briefs, reviewing outputs, updating escalation rules, retiring old agents. The remaining 40% is human-only work: escalations, planning, judgement calls.
The thing to notice: this role didn't exist three years ago. It's now one of the most leveraged seats in the company. Every founder running serious AI work is going to need someone in this seat. Most haven't named it yet.
If you're a founder and your week already looks like 60% briefing and reviewing AI work, you are the hybrid. That's the seat to name and hand off. Not the next sales hire. Not the next engineer. This one first.
Example: MAP A HYBRID OPERATOR'S WEEK
You tell Claude: "Map the operations lead's week across briefing agents, reviewing outputs, retiring agents, and human-only work. Show the percentage split and the leverage profile."
WHAT THIS PATTERN PRODUCES
- The most leveraged seat in a small AI company: one human, many compounding outputs.
- Clarifies the next hire. It's often not "another senior." It's the operator that runs the agents.
- Makes operational quality visible. Agent management becomes a discipline, not a side task.
WHERE IT BREAKS
- The seat doesn't exist in most companies yet. No obvious resume signal to filter for.
- Burnout risk if the operator owns too many agents without retiring the worst.
- Hard to backfill. Context on each agent lives in one head until specs catch up.
Use this when: your own week is mostly briefing and reviewing agents. Name the seat, write the JD, hand it off. The leverage shows up within a quarter.
PATTERN 03 · REPORTING LINES
The chart looks like a network, not a tree.
Lines go up, sideways, and down.
In a traditional company, reporting lines go up. Junior reports to senior. Senior reports to lead. In an AI company, reporting lines also go sideways, to the agent the human manages, and down to the outputs the agent produces.
A line lead doesn't only have humans reporting to them. They also have agents producing on their behalf, and they're accountable for both. The chart looks more like a network than a tree. Up, sideways, down, all at once.
What this changes: performance reviews now include agent oversight. "What did your agents produce, and what did you do about the failures?" becomes a real review question. Output quality is a shared metric between human and agent, with the human accountable.
Example: REDRAW THE TEAM CHART WITH ALL THREE NODE TYPES
You tell Claude: "Take my current team of 5 humans. Add the 7 agents we run. Show me the new reporting structure with lines going up, sideways to agents, and down to outputs."
WHAT THIS PATTERN PRODUCES
- Every agent has one owner. No more orphaned outputs.
- Accountability becomes traceable: broken output → broken agent → traceable to a human.
- The chart shape matches the work shape. Networked, not hierarchical.
WHERE IT BREAKS
- A network chart is harder to print at A4 than a tree.
- Agents shared across two managers create ambiguity. Pick one owner.
- Redrawing the chart every quarter is real work. But the work pays for itself.
Use this when: you can't list who owns which agent off the top of your head. That's the moment to redraw.
The four patterns at a glance
How to start.
Pick the smallest version of this you can ship this week. List five agents your team is already running. Give each one a single human owner. Add them to the chart you actually share with new hires. That's pattern 02, applied at the smallest scale. Most of the leverage shows up immediately.
Don't try to redraw the whole structure on day one. The two rules do the heavy lifting: every agent has a manager, no human does an agent's job. Apply them once a week to the agents you can see. New ones surface as you go.
The companies that figure this out get a permanent leverage advantage. The ones that don't get a stack of orphaned tools and a team that's quietly burning out managing them informally.
The chart isn't a vanity diagram. It's the source of truth for who owns what. When agents enter the work, they need to be on the chart, or they don't really exist as part of your company.
Now do this.
Open a blank doc. On the left, sketch your team's current org chart, humans only. The version you'd show a new hire on day one.
On the right, list every agent your team is actually running. CRM autoresponders. Drafting assistants. Scheduling bots. Monitoring scripts. Any piece of software that produces work without a human pressing go. Most teams find five to twelve they hadn't counted as part of the company.
For each agent on that right-hand list, write one human's name as its manager. The ones with no name beside them are the orphaned outputs. That's the unfinished work.