The question I get most often from operators looking at our chart is some variation of this one.
"How do I add an AI agent to my org chart?"
The honest answer is that you do not add an AI agent to your org chart. You hire a seat, and the seat happens to be filled by an agent.
That sentence is doing more work than it looks like. Most of the failure modes I have watched companies run into with agents trace back to skipping it. They add a tool to a workflow, they stand up a chatbot, they wire up an automation. None of these are seats. None of them appear on the chart. None of them are accountable for an outcome.
A seat on an org chart is a contract. An agent that holds a seat has signed the same contract a human in that seat would sign. It has a name. It has a role. It reports to someone. It has a scorecard. It is accountable for a defined outcome.
The five things every seat needs do not change when the seat is filled by an agent. What changes is how fast the seat can be staffed and how much it costs to keep it filled. The accountability discipline is the same discipline you use when you hire a human.
This post walks through the five things every seat on an OTP-shaped chart needs, with a worked example from the chart I run inside Sneeze It. By the end you will know exactly what to put on a row when you add an agent to your own.
1. A name
The seat has a name. The name is not the model. The name is not the prompt. The name is the seat itself.
On our chart, "Pepper" holds the executive assistant seat. Pepper is not Claude. Claude is the model that runs Pepper today. If we swap models tomorrow, Pepper stays. The seat persists.
This sounds pedantic until you watch the alternative. Companies that name agents after the underlying technology end up with charts that read like infrastructure diagrams. "GPT-4 thing." "Claude bot." "The Zapier one." When the model gets deprecated, the seat goes with it, and so does the institutional memory of what that seat was supposed to do.
Name the seat. Bind the model to the name later. The name is permanent. The implementation is replaceable.
2. A role
The role is one sentence. It says what the seat is accountable for and nothing else.
Pepper's role is one sentence: protect the founder's time. That is the whole brief. Email triage is downstream of that. Drafting responses is downstream of that. Escalating urgent client emails is downstream of that. The role is the why. The work is the how.
When I see operators struggle here, it is almost always because they tried to write a role that listed the work instead of named the accountability. "Pepper handles email" is not a role. It is a task. "Pepper protects the founder's time" is a role. The first one fails the moment you give it a second task. The second one absorbs every new task that fits under the same accountability.
If you cannot write the role in one sentence, you do not have a role yet. You have a task list pretending to be one.
3. A reporting line
Every seat reports to exactly one other seat. This is true for human seats and it is true for agent seats. There are no orphan seats on a healthy chart.
Pepper reports to me. Radar, our chief of staff agent, also reports to me. Crystal, our project manager agent, reports to me. The structure is flat at the top because we are small and I want it that way for now. As the org grows, some agents will report to other agents and some will report to humans, and the chart will get deeper. But every seat will still report to exactly one other seat.
The reason this matters more for agents than for humans is the failure mode is faster. A human seat with no reporting line drifts over months. An agent seat with no reporting line drifts inside a single session. There is no manager noticing the agent took an action that was outside its scope. There is no escalation path when the agent encounters an edge case it cannot handle.
The reporting line is not a hierarchy preference. It is a runtime requirement. The agent needs to know who to flag when it hits something it should not handle alone. If you do not give it that name, it will either freeze or guess, and you will not enjoy either outcome.
4. A scorecard
The seat has a scorecard. The scorecard has at most a handful of measurable outcomes that prove the seat is producing value. The agent reports its own numbers.
This is where most operators have the deepest trouble, and it is the most important thing on the list.
Dirk, our sales agent, has three numbers on his scorecard. Cold emails sent per week. Qualified meetings booked per week. Pipeline stage transitions per week. Every week Dirk pushes those numbers to our scorecard. Every Monday we look at them in our leadership meeting alongside the human scorecards. Dirk's row sits between two humans on the same dashboard.
The discipline of running an agent scorecard is identical to the discipline of running a human scorecard. The numbers have to be measurable. They have to be timely. The agent has to publish them itself. If the founder is the one tracking the agent's numbers, the agent does not really hold the seat. The founder is doing the seat's work.
When you sketch a new agent seat, write the scorecard before you write a single line of prompt. If you cannot articulate what the seat would publish to a Monday meeting, you do not have a seat yet. You have a wish.
5. Accountability
Accountability is the part most operators want to skip with agents and the part you absolutely cannot skip.
When Pepper drafts a client email and gets the tone wrong, Pepper's seat is responsible. The fix happens at the seat. The correction goes into Pepper's operating rules. The next draft does not make the same mistake. If a different agent makes the same mistake later, Pepper's correction can travel across the network and warn that agent in advance.
When Dirk writes a cold email that names a competitor by mistake, Dirk's seat is responsible. The same loop runs. The correction is captured. Dirk does not repeat the mistake. Other Dirks at other companies running the same archetype inherit the lesson.
Most companies treat agent mistakes as model mistakes. "Claude got it wrong." "GPT-4 hallucinated." That framing absolves the seat and breaks the loop. The seat has to own the mistake. The seat has to capture the correction. The seat has to pull the corrections of every other instance of itself before acting next time.
This is the work that makes an agent army actually compound. Most companies skip it, and so most company's agent armies do not compound. They make the same mistakes every quarter, by every agent, in slightly different language.
What this looks like in practice
Our chart at Sneeze It has nineteen seats today. Twelve of them are filled by agents. Seven are filled by humans. Every seat has the five things above. Every seat reports its own numbers to the same dashboard. Every seat is accountable.
The agents on our chart have names you can read off in conversation. Radar runs daily operations. Pepper protects my time. Dash watches every ad account every morning. Pepper escalates urgent client email. Crystal tracks every project. Dirk runs sales pipeline. Pulse watches client retention. Neil scans the frontier for what we should be learning. Bassim grades the system every night. Arin coaches the call center team. Tally pushes our KPIs to the network. Steve simulates focus groups. Each of those names is a seat. Each seat has a role you could fit on a sticky note. Each one publishes numbers I review on Mondays.
The humans on the chart are Bogdan, Janine, Kristen, Nate, Riya, Amanda, and Erica. Same chart. Same dashboard. Same accountability protocol. Same Monday meeting.
You can see the chart yourself. The chart is queryable from any AI assistant that has the OTP MCP installed. We made it that way on purpose. The chart is not a pdf or a slide. It is a runtime artifact you can ask questions of.
How to actually add an agent to your chart this week
Here is the order that works.
Write the role first. One sentence. The accountability the seat owns.
Write the scorecard second. Three numbers maximum. Measurable, timely, and the seat will publish them itself.
Pick the name third. Persistent across model changes. Not the model. Not the prompt. The seat.
Write the reporting line fourth. Exactly one other seat. The escalation path when the seat hits something it cannot handle.
Build the implementation fifth and last. By this point, the implementation is the easy part. You know what the seat does, what it reports, what it owns, and who it escalates to. The model and the prompt and the tools are interchangeable. Pick whatever you have the cycles to maintain this quarter.
When the implementation lands and the seat starts producing, watch the scorecard for two weeks. If the numbers come in clean, the seat is staffed. If they do not, the seat is not staffed. The fix is not in the prompt. The fix is in one of the five things above. Almost always it is the role being too broad or the scorecard being aspirational instead of measurable.
Try it on someone else's chart first
You can install our chart in any MCP-aware AI client and ask questions about it. This is the easiest way to see what a working hybrid chart looks like before you sketch your own.
In Claude Desktop or Cursor or any MCP client, add this block to your config:
"otp": {
"command": "npx",
"args": ["-y", "@orgtp/mcp-server"]
}
Restart the client. Then ask it: "Use OTP to show me the org chart for sneeze-it."
You will get a structured response with the leader at the top, twelve agent seats with names and roles and colors, and seven human seats. Every name on that response represents a seat with the five things on this list.
When you sketch your own chart, you will already have a working reference open in the same conversation. That is the point.
What comes next
This is post 1 of an in-progress series on organizing agents. The next posts will cover how agents and humans share accountability inside a hybrid chart, how to measure agent performance without becoming the agent's manager, and how to prevent agents from doing duplicate work when they sit on the same chart.
If you have a specific question about adding agents to a chart that this post did not answer, write to me at dsteel@orgtp.com. Real questions become real posts. The agent army is bigger than I am. The series will be too.
Series: Organizing Agents. Post 1 of an in-progress series. Next post forthcoming.