The things you still run that no longer serve you
You have agents — mental models, habits, routines, recurring commitments, automated behaviors — that were created for a context that no longer exists. The job changed. The relationship changed. The project ended. The tool was replaced. But the agent keeps running. You still check the dashboard. You still attend the meeting. You still organize your notes in a system designed for a problem you solved two years ago.
These are legacy agents. And the defining feature of a legacy agent is not that it is old. It is that no one — including you — can clearly articulate why it still runs.
In software engineering, Michael Feathers defined legacy code with a single criterion: code without tests. Not old code. Not bad code. Code that lacks the feedback mechanisms to tell you whether it still works as intended. The age is irrelevant. A system written last month with no tests is legacy. A system written twenty years ago with comprehensive tests and clear documentation is not.
The same principle applies to your epistemic agents. A habit you built a decade ago that still produces measurable value and that you can articulate the purpose of is not legacy — it is mature. A habit you built six months ago that you can no longer explain the purpose of, that produces no visible output, and that you continue out of inertia — that is a legacy agent. The test is not age. The test is whether you can answer, clearly and specifically, what this agent does for you right now.
Why legacy agents persist: the psychology of not removing things
The obvious question is: if these agents no longer serve you, why don't you just stop? The answer involves at least four interlocking psychological mechanisms, each of which makes removal feel harder than continuation — even when continuation is more expensive.
Status quo bias. Samuelson and Zeckhauser demonstrated in their 1988 study that people disproportionately prefer the current state of affairs, even when alternatives are objectively better. In experiments involving health plan choices and retirement programs, participants consistently chose whichever option was labeled as the "current" one — regardless of its actual merits. The bias is not about evaluating options and choosing the best one. It is about the act of choosing itself feeling risky. Doing nothing feels safe because nothing visibly changes. The costs of inaction are invisible; the costs of action are salient.
This is exactly how legacy agents survive. You don't actively decide to keep checking that dashboard every morning. You simply never decide to stop. And because the cost of checking it (a few minutes, a little attention) is diffuse and invisible, while the cost of removing it (what if I need it? what if I lose something?) is vivid and immediate, the dashboard wins by default. Every day.
Sunk cost reasoning. Arkes and Blumer's foundational 1985 research on the psychology of sunk cost showed that people increase their commitment to endeavors in proportion to their prior investment — even when the future return on that investment is zero. In their field study, theater patrons who paid more for season tickets attended more plays, regardless of how much they enjoyed them. The investment had already been made. The money was gone. But the feeling of waste was intolerable, so they kept going.
Your legacy agents carry sunk costs. You spent three weeks building that note-taking workflow. You spent a year developing that morning routine. You invested real cognitive effort designing that decision-making framework. The effort is gone — it cannot be recovered. But removing the agent feels like admitting the effort was wasted. So you keep running it, not because it produces value, but because stopping would force you to acknowledge that it doesn't.
Loss aversion. Kahneman and Tversky's prospect theory established that losses are felt roughly twice as intensely as equivalent gains. Removing a legacy agent is experienced as a loss — you lose the routine, the familiarity, the sense that you have a system. Adding a better replacement is experienced as a gain — but a gain that feels half as significant as the loss feels painful. The asymmetry means that even when the replacement is clearly superior, the emotional math discourages the switch.
Identity attachment. This is the mechanism that makes legacy agents particularly resistant in the epistemic domain. When an agent has been part of your routine long enough, it stops being something you do and becomes something you are. "I'm the person who journals every morning" is not a description of a habit — it is an identity claim. And challenging an identity claim triggers defensive responses that have nothing to do with whether the journaling still serves its original purpose. You built the agent to solve a problem. The problem is gone. But the agent has fused with your self-concept, and now removing it feels like removing a piece of yourself.
Cognitive behavioral therapy identifies this pattern precisely. Core schemas — deep beliefs about who you are and how the world works — generate automatic thoughts that selectively attend to confirming evidence and filter out disconfirming evidence. A legacy agent that has become part of your identity generates automatic defenses: "It's still useful, I just can't see how right now." "I'll need it eventually." "It's not hurting anything." These are not reasoned evaluations. They are the immune response of an identity protecting itself from revision.
The zombie process analogy
In operating systems, a zombie process is a process that has completed execution but still occupies an entry in the process table. The process is dead — it produces nothing, responds to nothing, serves no function. But it persists because its parent process never issued the system call that would clean it up. The zombie consumes a small amount of memory. One zombie is trivial. A hundred zombies degrade system performance. A thousand zombies crash the system.
Your legacy agents are cognitive zombies. Each one occupies a slot in your attention, your routine, your identity. Each one individually seems too small to matter. But they accumulate. Five legacy agents that each take ten minutes a day consume nearly an hour. Five legacy agents that each generate a small amount of decision fatigue — should I do this? is this still working? — compound into a background hum of cognitive load that you can't trace to any single source.
The zombie analogy is precise in another way: the zombie exists because the parent process failed to clean up. Your legacy agents exist because you — the parent process — never performed the equivalent of the wait() call. You never checked whether the agent had completed its purpose. You never issued the signal that would allow it to be removed from the table. The agent didn't refuse to die. You never asked it to.
Legacy agents in organizations: the trillion-dollar parallel
The pattern scales. Organizations carry legacy systems for the same psychological and structural reasons individuals carry legacy agents, but the costs are measured in billions.
Industry estimates place global technical debt at $1.52 trillion. CIOs report that 10-20% of budgets intended for new development are diverted to maintaining legacy systems. Eighty-five percent of business and technology leaders say that maintaining legacy systems hampers their ability to deploy new solutions. The systems persist not because anyone decided they should, but because no one decided they shouldn't.
Nelson and Winter's evolutionary theory of economic change explains the mechanism at the organizational level. Firms encode their knowledge in routines — standardized, repeatable patterns of behavior that persist through repetition. These routines are the organizational equivalent of habits: they coordinate action, reduce uncertainty, and encode tacit knowledge that no individual fully possesses. But the same stability that makes routines useful also makes them resistant to change. Routines persist even in the face of negative performance feedback, because changing a routine requires disrupting coordination patterns that everyone depends on. The cost of continuation is diffuse and distributed. The cost of change is concentrated and visible.
This is why legacy systems in organizations "begin to feel like part of the furniture," as one industry analysis puts it. The licensing costs are predictable. The impact on productivity goes unnoticed because it has always been there. The system becomes background — and backgrounds are, by definition, things you don't notice.
In machine learning operations, the same pattern appears with model drift. A model trained on historical data continues to make predictions even as the underlying data distribution shifts. The model doesn't announce that it has become unreliable. Its outputs gradually degrade. Without active monitoring — without the equivalent of Feathers' tests — the model continues running, consuming resources, and producing increasingly wrong answers that no one checks because the model is "already deployed."
The parallel to personal legacy agents is exact. Your habits, routines, and mental models were trained on historical data — the context of your life when you built them. That context shifts. The agents don't announce their obsolescence. They gradually degrade. And without active monitoring — without regular audits of what your agents are actually doing for you — they continue running, consuming cognitive resources, and producing increasingly misaligned outputs that you don't check because they are "already deployed."
The audit: how to find your legacy agents
Identification precedes retirement. You cannot remove what you haven't named. The audit has three phases.
Phase 1: Inventory. List every recurring agent in your life that runs at least weekly. This includes habits, routines, subscriptions, recurring meetings, automated workflows, maintenance tasks, information sources you check, and systems you maintain. Don't evaluate yet — just list. Most people undercount on the first pass. Go through your calendar, your phone's home screen, your browser tabs, your morning routine, and your evening routine. The inventory typically surfaces 30-60 items.
Phase 2: Articulation test. For each agent, answer two questions: (1) What specific outcome does this agent produce? (2) When did I last use that outcome to make a decision or take an action? If you cannot answer both questions clearly, mark the agent as a legacy candidate. The articulation test is the epistemic equivalent of Feathers' test criterion. An agent you cannot articulate the purpose of is an agent without tests — you have no way to verify that it is still working as intended.
Phase 3: Cost estimation. For each legacy candidate, estimate the weekly cost in three currencies: time (minutes spent), attention (does it fragment your focus or generate decision fatigue?), and opportunity (what could you do with the freed resources?). The time cost is usually the smallest. The attention cost is the most damaging. A legacy agent that takes two minutes but generates fifteen minutes of background anxiety costs seventeen minutes, not two.
The hardest part: acting on what the audit reveals
The audit will produce a list. The list will be uncomfortable. You will discover agents you have been running for years without clear purpose. And then the psychological machinery described earlier — status quo bias, sunk cost reasoning, loss aversion, identity attachment — will activate to prevent you from acting on what you found.
This is where the lesson connects back to the broader Phase 30 sequence. L-0588 (Agent retirement criteria) established the principles for deciding when an agent should be retired. L-0589 (Clean agent retirement) described how to retire agents without losing what they carried. L-0595 (Agent templates) showed how to create reusable patterns so that retirement doesn't mean starting from scratch. This lesson adds the missing piece: the agents that most need retirement are the ones you have the hardest time identifying, because they have become invisible through familiarity.
The agents you built deliberately and recently are easy to evaluate. The agents that have been part of your routine so long that they feel like part of you — those are the legacy agents. And the fact that they feel like part of you is precisely what makes them persist.
From identification to documentation
Naming your legacy agents creates an immediate practical problem: some of them contain embedded knowledge that will be lost if you simply delete them. The dashboard you built two years ago encodes assumptions about what metrics matter. The morning routine you no longer follow encodes a theory of productivity that might be partially right. The note-taking system you abandoned carries organizational logic that could inform your next system.
This is why the next lesson — L-0597, Agent documentation lifecycle — follows directly from this one. Before you retire a legacy agent, you need to extract what it knew. Documentation is the bridge between identification and clean removal. Without it, retirement becomes destruction. With it, retirement becomes composting — the old agent's knowledge feeds the new one.
The sequence is: audit, document, retire. Skip the documentation step and you will either refuse to retire agents (because you're afraid of losing what they contain) or retire them and lose valuable embedded knowledge. Neither outcome serves you.
Start the audit today. The legacy agents are already running. They have been running for longer than you think. The only question is whether you will notice them before they consume more than you can afford.