Not every agent works. Some seats turn out to be poorly defined. Some agents drift past prompt fixes. Some get absorbed into other agents' jobs. Some were the right idea at the wrong time. Retiring an agent is part of running an AI-integrated EOS® company, the same way moving on from a human team member is part of running any company.
What is different is that most leadership teams have never thought about how to fire an agent. They fire it badly. The team loses the lessons. The next agent inherits the same problems. The agent layer accumulates cruft.
A retirement protocol fixes this. We use one. It came out of the experience of retiring our first agent, Jeff, and discovering that the act of retiring him produced more lessons than the act of building him.
When an agent should be retired
Five signals.
Signal one: chronic scorecard miss. The agent has missed its weekly scorecard target for four consecutive weeks despite SOP updates. Not a one-off. A sustained pattern.
Signal two: persistent Core Value violation. The agent produces output that violates Core Values, and prompt fixes have not solved it. This is rare but serious. Retire fast.
Signal three: seat overlap. Another agent has absorbed the function. The agent in question no longer has a unique seat on the Accountability Chart. Keeping it creates confusion.
Signal four: dependency on a tool that disappeared. The agent was built to read from a system that the company no longer uses. Rebuilding it for the new system is more work than building a fresh agent.
Signal five: the Visionary or Integrator has lost confidence in it. Subjective, real, valid. If the dyad does not trust the agent, the agent is not contributing anyway. Retire it.
Any one of these signals is enough to start the retirement protocol. Two or more, especially if signal five is one of them, makes retirement urgent.
The protocol
Six steps. The protocol takes about a week if the team is ready and a few weeks if there is hesitation.
Step one: name the underperformance specifically. The Integrator writes a one-paragraph case for retirement. Which signals are firing. With evidence. This is the agent equivalent of a termination memo. Specific, fact-based, calm.
Step two: hold the hearing. Yes, a hearing. The leadership team gathers (or the Visionary-Integrator dyad if smaller). The agent's accountability partner presents the case. Other team members can argue for keeping the agent if they see something the case missed. This is not theater. The hearing prevents quiet retirements where the lessons go undocumented.
Step three: ask the agent. Hand the agent its own case for retirement and ask it to argue back. Specifically ask: "Defend your continued existence in this seat. What did you do well. What did you miss. Should you be retired." The model's response often contains insight the team did not have. Sometimes the agent argues persuasively for a focused continuation. Sometimes the agent recommends its own retirement.
This step matters. It is the difference between a clean retirement and a sloppy one. The agent's self-assessment is part of the record.
Step four: make the decision. The decision is human. The leadership team or the dyad votes. Retire or continue. If the decision is to continue, document the conditions under which retirement would proceed, with a date check-in.
Step five: document the lessons. Write a one-page retirement memo. What was the seat supposed to do. Why was it the wrong seat. Which other agents absorbed the function. What would we do differently if we built this seat again. Add to the company's Issues archive.
Step six: shut down cleanly. Disable the agent's schedule. Remove the agent from the Accountability Chart. Update any other agents that referenced this one. Archive the SOP. Notify the team.
The protocol is calm and dignified by design. The agent is not a person. The protocol is also not for the agent. The protocol is for the company, because the act of retiring well preserves the institutional lesson.
What we learned from Jeff
Jeff was our analyst agent. We retired him formally in April 2026. The hearing went exactly as the protocol describes.
Jeff's self-defense was the most useful part. We asked him to argue for his continued existence. He listed his accomplishments honestly. He also listed his failures honestly. He explicitly cited a soul-document line about creators sometimes needing to let go. He recommended his own retirement.
The team accepted. We wrote the retirement memo. We disabled Jeff's schedule. We updated Dash and Pulse and Dirk to absorb his remaining missions. We added the lesson to our Issues archive.
The lesson: agents need sharply defined seats. Jeff's seat was three different jobs that should have been three different agents. We held onto Jeff longer than we should have because the seat overlap was real but unacknowledged.
The single artifact from Jeff's retirement that has shaped how we build now: every new agent gets a clear single-seat job description and a sentence describing which existing agents it does not duplicate. Before Jeff, we did not check for duplication.
This is what a well-run agent retirement produces. A specific procedural improvement that makes the next agent better. Without the hearing, this lesson would not exist.
Why a hearing
Two reasons people resist the hearing format.
"It feels theatrical." Sometimes. The format also forces seriousness. The team is on the hook to argue both sides. The agent is on the hook to defend itself. The proceedings get documented. Quiet retirements never produce these artifacts.
"It feels like we are anthropomorphizing the agent." A little. The format also acknowledges that the company has invested real time, money, and institutional trust in this seat. Treating the retirement as a decision worth a serious meeting respects the investment.
Companies that retire agents informally lose the lessons. Companies that retire agents formally compound the lessons. Choose accordingly.
What about agents that just stop working
Some agents fail not from underperformance but from external causes. A vendor outage. A model deprecation. A platform sunsetting. These are different from performance retirements.
For external-cause retirements, run a shortened version of the protocol. Skip the hearing. Document the cause, the impact, and the replacement plan. The agent did not fail, the substrate failed. The lesson is about substrate, not seat.
What to tell the human team
When an agent retires, the team should hear about it. Two reasons.
Reason one: psychological hygiene. Team members who relied on the agent need to know it is gone. Otherwise they refer to outputs that no longer exist.
Reason two: institutional learning. The lessons from the retirement are useful to the team for future agent decisions. Spread them.
The format we use: a one-paragraph note in the next L10® under Customer or Employee Headlines (depending on the seat). "Jeff has been retired. Functions absorbed by Dash, Pulse, and Dirk. Lessons in the Issues archive. Net effect on the team's workflow: minimal." Move on.
FAQ
What if the agent disagrees with retirement? The decision is the human team's. The agent's argument is input. If the agent argues persuasively that retirement is wrong, listen carefully, then decide.
Should the agent's soul or persona be archived? Optional. Some teams keep an archive of retired agents' job descriptions and key outputs as institutional memory. Others delete. Either is defensible.
What about reusing the agent's name? Tempting but confusing. Use a new name for a new agent. Avoid the implication that the new agent is continuous with the retired one.
What if we want to rehire the agent later? Document the conditions under which we would. New SOP, new seat, new model, new evidence base. Treat the rehire like a new hire.
EOS®, Entrepreneurial Operating System®, V/TO™, Level 10 Meeting®, L10®, Rocks™, Scorecard, Issues List, Accountability Chart, Core Values, Customer Headlines, Employee Headlines, Visionary, and Integrator are concepts and trademarks of EOS Worldwide, LLC. This article is an independent practitioner perspective and is not affiliated with or endorsed by EOS Worldwide.