Gartner published a six-step framework for managing AI agent sprawl. They called agent sprawl "the new Shadow IT." That is good market validation. It means the problem is real, it is named, and it is being watched at the board level.
It does not mean anyone has solved it.
A framework is advice with a structure. An operating system is a thing that runs. The CIO community has plenty of the first and almost none of the second. That gap is the thing I want to be precise about, because it keeps getting obscured by how the advice is packaged.
Here are five differences that matter.
1. A framework describes seats. An operating system fills them.
Gartner's six steps include "centralized agent inventory" and "agent identity, permissions, and lifecycle." Those are correct. You should have an inventory. You should have lifecycle discipline. You should know which agents are running, what they own, and how you retire them when the role changes.
But describing those things and doing them are different problems.
At Sneeze It we run roughly ten agents on a single org chart right now. Each seat has one owner, one set of metrics, and one place on the Monday scorecard. Radar runs the daily brief and calendar. Tally pushes KPI values from local sources to the chart four times a day. Dash reads ad spend and call center data across 39 accounts and flags when something breaks pattern. Nick runs cold prospecting in health and wellness. Each seat is named. Each seat is filled. Each seat has a row on the same dashboard the humans are on.
Filling those seats did not happen because we read a framework. It happened because we made specific decisions, in sequence, about role design, measurement, coordination, and what to do when a seat stops earning its place.
A framework can prompt those decisions. It cannot make them for you.
2. A framework documents the problem. An operating system creates accountability.
The advisory side is ahead of the academy on naming the problem. Gartner, as reported by CIO.com, says roughly 50 or more agents are already running inside large enterprises, and agent sprawl has become the new Shadow IT. The Deloitte State of AI in the Enterprise 2026 (n=3,235) found that only 21% of companies have a mature governance model for agentic AI. That means roughly 80% of companies deploying agents have no real governance at all.
Those numbers are useful for making the case internally. They are not the same thing as accountability.
Accountability requires a name on a row, a number in a cell, and a conversation on a specific day about what happened. Without those three things, governance is a policy document that lives in a shared drive and gets cited in a board deck. It does not change what the agents do Monday morning.
Our agents are accountable the same way our humans are accountable. Bogdan, our COO, has rows on the scorecard. So does Janine. So does Kristen. So does Dirk, our sales agent, who tracks cold emails sent, qualified meetings booked, and pipeline stage transitions. When Dirk's row drops, the conversation is the same conversation you would have about any seat. What changed. What is the fix. When is it recovered.
The governance is not a policy. It is a weekly row review.
3. A framework ends at design. An operating system includes retirement.
This one is easy to miss because it does not show up in the framework until you need it.
Every agent fleet accumulates seats that outlive their purpose. The role changes, the work moves, a better agent upstream absorbs the function. If you have a framework but no retirement discipline, those seats just sit. They consume compute. They emit signals nobody reads. They create the impression that the fleet is larger and more coordinated than it actually is.
We retired Jeff, our data integrity agent, in April. It was the first agent retirement at Sneeze It. We held a formal process. Jeff was asked to account for what his seat had produced and what had changed. He assessed honestly that his core functions had migrated to other seats (Dash absorbed the ad pacing and account monitoring work; Dan absorbed the data architecture role). We redistributed the capabilities and closed the seat.
That process is not described in any framework I have read. Most frameworks end at "retiring redundant agents" as a bullet under lifecycle management. What they do not include is the conversation discipline, the redistributed accountability, and the explicit record of what the seat produced before it closed.
An operating system has to include retirement because the fleet is always changing. Advice about lifecycle is not the same as a lifecycle.
4. A framework is static. An operating system learns.
This is the hardest thing to build and the easiest thing to omit from a framework, because frameworks are written once.
When a human agent (manager, analyst, setter) makes a mistake, the correction travels. The manager tells the analyst what went wrong. The analyst adjusts. The team lead sees the correction happen and the pattern propagates through the people on the team.
With AI agents, corrections tend to evaporate. The agent is corrected once, the session ends, and the next session starts without the correction baked in. The fleet does not get smarter. The same mistake recurs.
We run a correction protocol at Sneeze It. When David corrects any agent, the correction is captured as a structured learning and stored where every agent can read it before the next task. Neil, our learning agent, runs a frontier scan every week and gates what survives against two questions: will this make the fleet better, and is what we already have better. Only verified improvements get through. Bassim evaluates the fleet's maturity level and identifies where the next bottleneck is.
That is a learning loop, not a framework. A framework describes what a learning loop should do. An operating system runs one.
5. A framework names the gap. An operating system closes it.
MIT Sloan's research arm, MIT CISR, has published papers on "digital colleagues," decision rights for autonomous agents, and governing multiagent systems. That research is genuinely ahead of the teaching. CMU's Heinz College runs the LEAAID certificate, which goes further than any other school I can verify, with hands-on agent building and modules on agentic governance and accountability. These are real contributions.
But even CMU, the closest any school gets, teaches you to build and govern one agentic capability. It does not teach you to run the fleet. "Agent fleet," "agent sprawl," "fleet health," "who owns the agent stack" show up nowhere in any verified curriculum.
The gap is not awareness. The gap is not even frameworks. The gap is a live operating system: a chart where humans and agents hold named seats with one-seat-one-owner, scorecards that update continuously, lifecycle that includes real retirement, and a coordination network that captures what the fleet learns.
That is the gap OTP fills. Not advice. Not a framework. A running system.
The goal is simple to state and hard to build. Let agents carry the operational work, so people are free for the work that matters. Getting there requires more than knowing what to do. It requires a structure that runs the discipline every day, whether you are watching or not.
A framework tells you what to do. An operating system is what you do every day.
See the live chart
From any MCP client with the OTP server installed, you can query the Sneeze It agent fleet, pull the seats, and see which agents hold active roles, which have been retired, and what metrics each seat owns.
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 list the agent seats on the Sneeze It org chart, including any retired seats, and show what each seat is accountable for."
What comes back is not a framework. It is the actual structure, running live, with names and accountabilities attached. That is what a running operating system looks like from the outside.
Series: AI CIO. Part 13 of an in-progress series.