The agency that existed only because no one had permission to kill it
In 1976, Colorado became the first U.S. state to enact a broad sunset law requiring most government agencies to undergo periodic review and justify their continued existence. If an agency could not demonstrate its effectiveness and necessity, it would be automatically abolished. The law was not created because Colorado had a surplus of visibly failing agencies. It was created because the state had accumulated agencies that no one could remember creating, that served purposes that no longer existed, and that consumed budgets simply because they had consumed budgets last year. The absence of formal retirement criteria meant that every agency, once created, was effectively immortal. Termination required an active decision, and no one had the incentive to make it.
Texas followed with its own sunset review process, and the pattern it uncovered was consistent: agencies that had outlived their purpose did not announce this fact. They continued to operate, continued to file reports, continued to request funding. They exhibited all the surface indicators of functioning organizations while producing little or no value aligned with their original mandate. The sunset review forced a question that had never been asked: not "is this agency doing something?" but "is what this agency does still worth doing?"
This is exactly the problem you face with your cognitive agents. You have habits, routines, review protocols, and decision frameworks that you built to solve specific problems in specific contexts. Some of those problems have been solved. Some of those contexts no longer exist. But without explicit retirement criteria — without a sunset clause that forces the question — those agents will persist indefinitely, consuming attention, occupying time slots, and crowding out agents that actually serve your current reality.
Why agents resist retirement
The research on why people fail to quit things that are no longer working converges on three mechanisms, each of which operates powerfully on cognitive agents.
The sunk cost trap. Hal Arkes and Catherine Blumer's foundational 1985 research demonstrated that people continue investing in failing ventures specifically because of what they have already invested, not because of what they expect to gain. The more you have invested in building, tuning, and running an agent, the harder it becomes to retire it — even when the evidence clearly shows it has stopped producing value. As Annie Duke argues in her research on strategic quitting, humans are not merely reluctant to quit; they actively escalate commitment in the face of negative feedback, increasing investment precisely when they should be withdrawing. The sunk cost of the time you spent building a morning review template, the effort you invested in a journaling protocol, the weeks you spent calibrating a decision matrix — these past costs have zero bearing on whether the agent serves you today, but they exert enormous gravitational pull on your willingness to let it go.
Identity attachment. You do not just run your agents. You identify with them. "I am the kind of person who does a weekly review." "I am someone who tracks my habits." "I use the Eisenhower Matrix because I am systematic." When an agent becomes part of your self-concept, retiring it feels like retiring a piece of who you are. This is not an exaggeration of the psychology — research on identity-based habits shows that habits become woven into self-narrative, and discontinuing them triggers the same discomfort as other identity threats. The agent is no longer a tool you use; it is a trait you possess. And you do not voluntarily give up traits.
The absence of a forcing function. Colorado's sunset law worked because it reversed the default. Before the law, agencies continued unless someone actively terminated them. After the law, agencies terminated unless someone actively renewed them. The procedural shift was everything. Without a sunset clause, the default for your cognitive agents is persistence — they continue unless you make an active decision to stop them. And active decisions are expensive. They require attention, deliberation, and the willingness to overcome the sunk cost and identity pressures described above. Most days, it is easier to just keep running the agent, even when you suspect it has outlived its purpose.
These three forces — sunk cost, identity, and the absence of a forcing function — create a systematic bias toward agent accumulation. Over months and years, you build new agents but rarely retire old ones. The result is agent sprawl: a growing portfolio of routines and systems competing for finite attention, many of which no longer justify the resources they consume.
Kill criteria: the pre-commitment that makes retirement possible
Annie Duke's research on strategic quitting identifies one mechanism that reliably overcomes the bias against quitting: kill criteria established in advance. Kill criteria are specific, measurable conditions defined before you need them — conditions under which you commit to retiring an agent regardless of how you feel about it in the moment.
The principle is borrowed from financial trading, where stop-loss orders work precisely because they are set before the emotional pressure of a losing position distorts judgment. A trader who sets a stop-loss at 15% below purchase price does not have to decide in real time whether to sell a declining stock. The decision was made in advance, when cognitive clarity was high and emotional stakes were low. The stop-loss executes automatically, bypassing the sunk cost escalation that would otherwise keep the trader holding a losing position.
Kill criteria work the same way for cognitive agents. You define them when you create the agent — when you are thinking clearly about what the agent is for and what would make it no longer worth running. Then, when conditions change and the criteria are met, the retirement decision is already made. You are not deciding whether to quit. You are honoring a commitment you made to yourself when you could think straight.
The best kill criteria combine three elements: a measurable state, a time boundary, and a decision. "If this habit has not produced a measurable improvement in [specific outcome] within 90 days, I will retire it." "If I skip this review more than three times in a month without noticing any consequence, it is retired." "If the context this agent was built for changes — new job, new role, new living situation — the agent enters a 30-day review period and must re-earn its place."
Without kill criteria, retirement is a judgment call made under the worst possible conditions: when you are emotionally invested, when the sunk costs are salient, and when the path of least resistance is to keep the agent running. With kill criteria, retirement is a commitment honored under the conditions you chose in advance.
Five categories of retirement criteria
Drawing from product end-of-life frameworks, ML model deprecation practices, and organizational sunset reviews, retirement criteria for cognitive agents fall into five categories. An agent should be evaluated for retirement when any one of these criteria is met.
1. Performance degradation
The agent's output has declined below the threshold that justified its existence. This is the most intuitive criterion and the one most people think of first — but it is also the hardest to measure without explicit metrics.
In machine learning, models are retired when concept drift causes prediction accuracy to drop below acceptable levels. The key insight from ML practice is that performance degradation is measured against a baseline, not against a feeling. You do not retire a model because it "seems less accurate." You retire it because its F1 score dropped below 0.8, or its false positive rate exceeded 5%, or its predictions diverged from ground truth by more than a defined margin.
For cognitive agents, this means you need a performance baseline. What did this agent produce when it was working well? How much time did it save? What decisions did it improve? What outcomes did it enable? If you cannot answer these questions, you cannot detect performance degradation — which means you cannot apply this criterion, which means you will hold onto degraded agents indefinitely.
Operationalize it: When you create an agent, document what "working well" looks like in measurable terms. During periodic reviews, compare current performance against that baseline. If performance has fallen below 70% of baseline for three consecutive review periods, the agent meets this retirement criterion.
2. Context obsolescence
The world the agent was designed for no longer exists. This is the concept drift problem from machine learning applied to life circumstances: your role changed, your priorities shifted, your environment is different, the problem the agent was built to solve has been solved or has transformed into a different problem.
Product managers use the maintain-refresh-retire framework specifically to address this: when the market a product was built for has shifted, the product enters evaluation for retirement regardless of whether it still technically functions. A product that works perfectly but serves a market that no longer exists is not a working product. It is a legacy artifact consuming resources.
Operationalize it: Every agent should have a documented context — the specific circumstances, role, goals, and environment it was designed to serve. When that context changes significantly (new job, new role, completed project, life transition), every agent built for the old context enters a 30-day review period. The agent must demonstrate relevance to the new context or be retired.
3. Cost-value inversion
The maintenance cost of running the agent exceeds the value it produces. This is the criterion that catches agents that still technically work but are no longer worth the resources they consume.
Gartner's research on legacy systems found that by 2025, organizations were spending 40% of IT budgets on maintaining technical debt — systems that still functioned but consumed resources disproportionate to their value. The same dynamic applies to personal cognitive infrastructure. A weekly review that takes 90 minutes but produces 10 minutes of actionable insight has a cost-value ratio that no longer justifies its existence. A habit tracker that takes 15 minutes daily but has not changed your behavior in three months is consuming time for nothing.
Operationalize it: Periodically estimate the time and attention each agent consumes (its cost) and the decisions, actions, or outcomes it enables (its value). If the cost exceeds the value for two consecutive review periods, the agent meets this retirement criterion.
4. Redundancy
Another agent, system, or capability now serves the same function. This happens naturally as your cognitive infrastructure matures — you build better systems that subsume the work of earlier ones. The earlier agents do not fail; they become redundant. But without explicit redundancy checks, both the old and new agents continue running, creating unnecessary overhead and potential conflicts.
In software architecture, this is the legacy system decommissioning problem. The new system is live and handling the workload, but the old system keeps running because no one issued the decommission order. Data flows through both systems. Teams maintain both. Resources are split. The solution is explicit: when a new system covers the functionality of an old one, the old one enters a decommission timeline with a hard end date.
Operationalize it: When you create a new agent, explicitly check whether it subsumes any existing agent's function. If it does, the old agent enters a 30-day phase-out period. If the new agent performs adequately during that period, the old agent is retired.
5. Sustained non-use
You have stopped running the agent, and nothing bad has happened. This is the simplest and most empirically honest criterion. If you have skipped an agent repeatedly without negative consequences, the agent is telling you — through your own behavior — that it is not needed.
Research on habit discontinuation shows that habits weaken when their contextual cues are disrupted. But the inverse is also informative: when a habit stops firing and nothing degrades, the habit was either redundant or addressing a problem that no longer exists. Your non-use is data. Treating it as laziness or failure is a misread. Often, it is the most honest signal available about an agent's relevance.
Operationalize it: If an agent has not fired for three consecutive scheduled occurrences, or if you have actively skipped it three times without noticing a downstream consequence, the agent meets this retirement criterion.
The sunset clause: reversing the default
All five criteria share a structural problem: they require you to actively evaluate each agent against its criteria, which means they are themselves vulnerable to neglect. The solution is borrowed directly from sunset legislation: reverse the default so that agents expire unless actively renewed.
A sunset clause for a cognitive agent works like this: when you create the agent, you assign it a review date — 30, 60, or 90 days from creation. On that date, the agent is considered expired unless you explicitly renew it. Renewal requires a brief written justification: what value has this agent produced in the last period, and why should it continue? If you cannot write that justification, or if you do not bother to write it, the agent is retired.
This is the Colorado model applied to personal infrastructure. The procedural reversal is what makes it work. You are no longer asking "should I retire this agent?" — a question that triggers sunk cost bias, identity attachment, and status quo preference. You are asking "should I renew this agent?" — a question that requires evidence, not inertia.
The practical implementation is simple. Maintain a list of your active agents with their next review dates. When a review date arrives, you have three options: renew with justification, retire, or replace with something better. If you take no action, the agent is retired by default. This forces a minimum level of active engagement with your portfolio, prevents the accumulation of zombie agents, and makes retirement the path of least resistance rather than the path of most.
The emotional skill of letting go
None of these frameworks eliminate the emotional difficulty of retiring an agent. Kill criteria, sunset clauses, and the five retirement categories are cognitive scaffolding — they make the right decision legible. But you still have to execute it. And execution means sitting with the discomfort of admitting that something you built, something you invested in, something you identified with, has reached the end of its useful life.
This is a learnable skill. Annie Duke's research frames quitting as a competency, not a character trait. The more you practice deliberate retirement — naming the criteria, checking them honestly, executing the retirement — the lower the emotional barrier becomes. Each successful retirement builds evidence that letting go of a dead agent does not damage you. It frees resources. It clarifies your portfolio. It makes room for agents that actually serve the person you are now, not the person you were when you built the old one.
The goal is not to become ruthless about retirement. It is to become honest about it. Every agent in your portfolio should be there because it earns its place, not because you lack the criteria or the courage to ask whether it still does.
From criteria to process: bridging to clean retirement
Knowing when to retire an agent is half the problem. The other half is how. Retirement criteria tell you that an agent should go. But agents do not exist in isolation — they connect to other agents, feed into workflows, and occupy slots in your schedule and attention. Pulling an agent out of a running system without updating the dependencies creates gaps, confusion, and cascading failures.
This is why L-0589 — Clean agent retirement — follows directly from this lesson. Once you have criteria that identify an agent for retirement, you need a process that retires it without breaking everything that depended on it. The criteria give you the "when." The next lesson gives you the "how."