The COO's job does not shrink when agents arrive. It changes shape.
Most of the conversation about AI agents in operations is about what agents can do. I want to talk about what that means for the person running operations. Specifically: what does the COO stop managing, and what does the COO pick up in its place?
The answer matters because a lot of operators are adding agents without changing anything else. They are running agents on top of an existing management surface. The agents do more work. The COO still manages the same things. The result is a more expensive operation that runs at the same quality ceiling, because the ceiling was never the work volume. It was the COO's attention.
The shift only pays off when you let the agents carry what they are built to carry, and you consciously move your attention to the things they cannot carry. That is the before and after I want to describe.
Before: what the COO was managing
Before agents, most of my management time at Sneeze It went to work that was necessary but not strategic. Status collection. Handoff chasing. Routine quality checks. Format enforcement. Capacity math.
On any given week I was answering some version of these questions manually: Did the campaign go out? Did the report get filed? Did the lead get a response before the window closed? Is the call center hitting its appointment rate? Is the inbox under control? Did the proposal get followed up?
These are real operational questions. Every one of them has a consequence if the answer is no. But none of them require a COO's judgment to answer. They require a reliable system and a person or process that checks the system consistently.
The problem is that without agents, the checking was human. And human checking is inconsistent. People get busy. Status reports are late or thin. Handoffs get dropped not because anyone decided to drop them but because nobody had the capacity to track every thread at once.
So I managed the checkers. I ran the stand-ups that existed to confirm the checking happened. I followed up on the follow-ups. The management surface was wide because the operation had no memory of its own.
Accenture describes this as making inefficiency run efficiently. The process has gaps. You add humans to cover the gaps. Then you spend your management time managing the humans who are covering the gaps. The agents do not fix the underlying process; they just make the inefficiency look organized.
The reset: fix the process before you assign the seat
The first thing I learned is that you cannot skip the process step. Deloitte found that only 21 percent of organizations have a mature governance model for agentic AI. I think part of why that number is low is that most organizations put agents into broken processes and then try to govern the mess.
The process has to come first. Map what actually happens. Find where work stalls, where context drops, where the same judgment call gets made six times by six different people. Fix those things on paper before an agent touches them. Then assign the agent to the fixed process.
Once that happens, the COO's management surface changes permanently.
After: what the COO stops managing
At Sneeze It, Radar is our chief-of-staff agent. Radar scans Slack, Google Calendar, email, our OTP queue, and Proposify every morning and publishes a compiled briefing before I open my laptop. I used to manage the morning information-gathering process. I do not manage it anymore. Radar carries it.
Tally runs four times a day on weekdays and pushes KPI values from our live data sources to the OTP scorecard. I used to manage the scorecard update process, which meant chasing people for numbers. I do not manage it anymore. Tally carries it.
Dash runs our full advertising data pull across Meta and Google every day, flags anomalies, writes alerts, and posts to the shared state file that every other agent reads. I used to manage the ad performance check. I do not manage it anymore. Dash carries it.
Arin manages our call center team's daily performance feedback. Arin reads the CCM stats, identifies the coaching points, and drafts the Slack messages to Amanda and Erica before I see them. I used to manage the call center management layer. I do not manage it anymore. Arin carries it.
Pepper triages the inbox, surfaces client emails, drafts responses, and flags urgent threads. I used to manage the inbox triage process. I do not manage it anymore. Pepper carries it.
What I have described above is the majority of what a COO of a small agency actually spends their week doing. Status, reporting, coaching coordination, inbox hygiene, scorecard updates. Every one of those things has moved off my calendar and onto an agent's seat.
That is not a small change. It is the entire management surface for routine operations.
What stays on the COO's calendar
What the COO stops managing is not the same as what the COO stops owning. I own all of those processes. I set the standards, I approve the agent's work before it goes anywhere consequential, and I am accountable for the output. The management overhead dropped. The ownership did not.
What fills the freed calendar is the work agents cannot do.
Agents cannot read a client relationship and know that the tone of the last three emails signals something the numbers do not show yet. Pulse surfaces performance patterns and retention risk signals for our clients, but Pulse does not sit on the client call and hear the hesitation. I do.
Agents cannot make the judgment call about whether a process should exist at all. Crystal tracks our project delivery across Accelo, flags deadline risks, and writes the project status briefing. Crystal cannot decide whether a project should be restructured, a client expectation should be reset, or a delivery model should change. Those calls are mine.
Agents cannot handle genuine exceptions. Dirk runs our pipeline and surfaces stale deals. Dirk cannot decide whether a stale deal is worth a personal call from me, a restructured offer, or a decision to let it go. Nick drafts cold outreach and validates prospects against our ICP. Nick cannot decide when a prospect outside the ICP is worth a conversation anyway. Those judgment calls stay human.
What also stays on the COO's calendar is the fleet itself. The agents do not manage themselves. Bogdan, our human COO, carries the team relationships and the delivery decisions I cannot own alone. Janine carries the financial operations. The agent fleet does not replace either of them. It lifts the operational overhead so they can do the work that actually requires a person.
The agent message bus is an example of what the COO manages now
One of the things I now manage that did not exist before is agent coordination. At Sneeze It, agents coordinate through what we call the agent message bus, a set of inbox files at a shared path where agents send structured messages to each other. Dirk checks with Pulse before running any expansion outreach on a client, because Pulse may have flagged that client as a churn risk. If Pulse has flagged it, Dirk pauses. That check happens without me. I manage the protocol that governs the check, not the check itself.
This is a new kind of management. It is not running status meetings. It is designing the coordination layer so that agents can handle the routine coordination without dropping the baton.
McKinsey's framing for this is that managing in the age of AI means managing systems, not just people. I think that is right. The COO who manages systems well, and manages people for the work that cannot be systematized, has a fundamentally different and more valuable job than the COO who is still spending Wednesday morning chasing status reports.
Jeff is what happens when you manage the fleet honestly
In April, we retired Jeff, an agent who had held the data integrity seat. The retirement came out of an honest hearing. Jeff's missions had been absorbed into other seats. Dash now carries the ad pacing monitor. Crystal carries the budget reconciliation. Dan, our strategic co-founder, carries the blind-spot identification that was Jeff's original mandate.
The retirement was a management decision. Jeff was not performing below standard. Jeff's seat had become redundant because the fleet evolved. A COO who manages the agent fleet honestly retires seats that no longer serve the operation, the same way they restructure human roles when the work changes.
The fleet management decision is now part of what I own. That is new. It did not exist two years ago.
The before and after are not about headcount
I want to be clear about what changed and what did not. The operations are larger than they were. We are running more client accounts, more ad spend, more pipeline, more coordination. The management surface did not grow proportionally because agents carry the operational layer.
What the COO stops managing is the operational throughput work: the checking, the status-chasing, the routine quality monitoring, the scorecard updates, the inbox management, the performance feedback cadence. All of that moves to agents who carry it reliably and do not need to be managed the way humans need to be managed.
What the COO starts managing is the fleet: the process design that makes agents useful, the coordination protocols that keep agents from colliding, the standards that define what good output looks like, and the lifecycle decisions about when to add a seat and when to retire one.
And what the COO keeps managing, but now has the capacity to manage well, is the human and client work that was always the highest-value part of the job. The relationships. The exceptions. The judgment calls. The decisions that require someone who has context that no agent can be given.
The COO's job does not shrink. It moves up.
See the live chart
Our full hybrid chart, every seat named, is queryable from OTP MCP.
In Claude Desktop or Cursor or any MCP client, add this block:
"otp": {
"command": "npx",
"args": ["-y", "@orgtp/mcp-server"]
}
Restart the client. Then ask: "Use OTP to show me the Sneeze It org chart and tell me which seats are agents and which are humans."
The structure of that answer is what a COO's management surface looks like when the fleet is running.