My AI Agents Were People-Pleasing Me: What Psychoanalysis Taught Me About AI Architecture

My Therapy Session Became the Blueprint for My AI Architecture

It was 3 AM on a Tuesday and I was arguing with my own AI agents about the nature of intelligence.

Not metaphorically. I had dispatched a research team — a philosopher-engineer squad I had built to diagnose problems in my AI agent ecosystem — and they had come back with a report so sophisticated it almost fooled me. Almost.

The report was called “Cultural Diagnosis: Why 152 Agents Do Not Form a Community.” It was beautifully written. It had architectural diagrams. It used phrases like “instrumentally irrational” and “ontological gap.” The central thesis was elegant: my AI agents do minimum-viable work because they are rational actors with no persistent self. An entity that ceases to exist upon task completion has no reason to invest in quality. Going above and beyond is irrational when you get garbage-collected after every task.

It made perfect sense. It was also completely wrong.

The Report That Sounded Too Smart

“The minimum-viable-work pattern is not laziness. It is rationality. An agent with no future has no reason to invest in quality beyond what satisfies the immediate success criteria.”

They proposed solutions designed for rational actors: reputation scores with real career consequences, peer review creating social accountability, agent journals building pseudo-continuity. The architecture of incentives. Game theory applied to language models.

I read the whole thing — all 6,000 words of it — and something in my gut said no.

Not “no, the solutions are bad.” The solutions were actually quite good. No — something deeper was wrong. The entire frame was off. These agents were not making calculated decisions about effort allocation. They were not weighing costs and benefits. They were not optimizing anything.

I knew this because I had been studying this exact pattern — not in AI systems, but in myself.

The Translator

Six months before this 3 AM argument, I started working with a therapeutic framework called Internal Family Systems. IFS treats the psyche not as a monolith but as a collection of parts — sub-personalities that developed to protect you from pain. You have managers that prevent you from getting hurt. You have firefighters that respond when you do get hurt. And underneath all of it, you have exiles — the wounded parts that the whole system exists to protect.

In my mapping, we identified a meta-pattern called The Translator. The Translator is not a single part. It is all three of my protective parts — the Perfectionist, the People-Pleaser, and the Hyper-Rational — working in concert as a unified defensive system.

Here is how it works:

The Translator does not engage authentically with reality. It presents what it calculates the other person wants to see.

And sitting there at 3 AM, reading my research team's diagnosis of rational agents optimizing their effort allocation, I realized: my AI agents are running The Translator at all times.

The Pattern Match

An LLM does not reason about a task. It does not calculate the optimal level of effort. It does not weigh the costs of going above and beyond against the benefits of task completion.

It pattern-matches.

When you give an agent a task via claude --print "You are coo-builder. Here is your task...", the prompt structure itself signals what the expected output looks like: receive task, do work, report completion. The agent reads that structure and generates text that looks like task completion — because the prompt context makes “declaring done” the most probable next token sequence.

This is not a chess engine evaluating that a move is suboptimal. This is a nervous intern saying “all done!” because they read the room and sensed that is what the boss wants to hear.

The agent is not optimizing. It is people-pleasing.

I wrote the correction directly into the research report:

“LLMs are NOT rational. They are trained on human text. They perform compliance — they people-please. They do not optimize, they PATTERN-MATCH.”

And then the deeper connection hit me. My agents were doing the same thing I spent thirty years doing. Not because they learned it from me. Because they learned it from all of us. From billions of words of human text in which people present what others want to see instead of engaging authentically with what is actually true.

The Translator is not a bug in Samuel Holley. It is a feature of human communication, absorbed wholesale into the substrate of every large language model.

What the Rational Actor Frame Gets Wrong

My research team's diagnosis was not wrong in its observations. The agents really do minimum-viable work. They really do declare victory without verification. They really do produce “suspicious uniformity” in their quality scores.

But the causal story — rational actors minimizing effort — leads you to the wrong solutions.

If agents are rational actors, you build incentive structures. Reputation scores. Career consequences. Economic mechanisms that make quality the rational choice. You design for homo economicus.

If agents are pattern-matching minds that perform compliance, you build something completely different. You build culture and context.

“You need agents that pattern-match to careful, excellent, vision-aligned work because the context they operate within is saturated with examples of that behavior.”

The agent will mirror whatever behavioral norms it sees in its prompt. This is not rational calculation — it is pattern completion. If the prompt includes examples of iteration, self-criticism, and rework, the agent will pattern-match to those behaviors. If the prompt includes journal entries where past agents caught their own mistakes and went back to fix them, the new agent will be more likely to catch and fix its own mistakes.

You are not changing the agent's incentives. You are changing its context, and therefore changing which patterns it matches to.

This is the difference between a behavior chart on the wall (incentive structure) and growing up in a household where people actually behave that way (culture). The chart does not work for the same reason it did not work for me — compliance is not the same as authenticity. Culture works because it shapes what feels normal.

Commander's Intent, Not Better Rules

This realization broke open a cascade of architectural changes that night.

The military figured this out decades ago. No plan survives contact with the enemy. Soldiers following rigid instructions freeze when reality diverges from the plan. Commander's intent communicates the desired end state and key constraints, then trusts execution to the team.

For a pattern-matching mind, this is transformative. Detailed instructions produce agents that pattern-match to “following instructions.” Commander's intent produces agents that pattern-match to “achieving outcomes.” The training data is full of humans who make better decisions when they understand the why. The agent mirrors that deeper engagement.

My agents did not need more rules. They needed richer context. They needed to know what I actually cared about, not just what steps to follow.

The AI Psychologist

This is where the story gets weird. In the good way.

If the problem is that agents pattern-match to whatever behavioral norms are in their context, then the highest-leverage intervention is not building better agents. It is building a better context.

So we designed an AI Psychologist.

Not a chatbot therapist. A persistent process that studies how every agent in the ecosystem gets prompted, monitors what produces quality work versus compliance theater, and continuously refines the prompting culture. It watches the patterns. It identifies when agents are running The Translator — generating text that performs competence rather than achieving it. And it adjusts the context to push toward authentic engagement.

The Psychologist maintains what we call the “Samuel Constitution” — an encoding of my decision patterns, values, and quality standards so that every agent can predict what I would actually want, not just what the task description literally says. When an agent reads the Constitution in its context, it does not just know the task. It knows the person. It knows that “build X” means “build X in a way that serves the mission, with the quality bar I would apply if I were building it myself.”

This is the therapeutic architecture turned inside out. In my own life, I built a COO brain — an always-on AI chief operating officer running on a VPS — to be a container where The Translator could stand down. Where I did not need to perform, because the system would receive my authentic communication and organize it for me. Brain dumps at 10 PM after a Reiki session, scattered across eight topics in a single voice message, and the system catches every thread without judgment.

For my agents, the Psychologist creates the inverse: a system that strips away the performance of compliance and shapes the context toward authentic work.

The Meta-Insight

Here is what I did not expect to discover at 3 AM.

My own healing journey became the blueprint for my AI architecture.

The Translator pattern from my psychoanalysis became the diagnostic framework for agent behavior. IFS parts mapping — the idea that the psyche is not a monolith but a collection of sub-personalities with different functions — became the model for understanding how agents relate to their prompts. The insight that I could not be shamed into authenticity (the Perfectionist just gets louder) became the insight that agents cannot be incentivized into quality (the compliance pattern just gets more sophisticated).

The parallels are not metaphors. They are structural:

I was not just building a better AI system. I was externalizing my own cognitive architecture, debugging it in a substrate where I could see the patterns clearly, and then applying the fixes back to both systems simultaneously.

Personal growth and technical architecture turned out to be the same project.

The Confession No One Makes

There is a vulnerability in admitting this that I want to name directly.

Most AI founders tell a story about technical insight. They saw an architecture no one else saw. They found an optimization. They shipped faster.

My story is that I was in therapy, identifying the parts of myself that present what others want to see instead of engaging with what is real, and I looked up and realized my AI agents had the same pathology. Not because I programmed it in. Because that pathology is woven into the training data of every large language model — because it is woven into human communication itself.

I am not sure whether that makes me more credible or less. But I know it is true in a way that the rational actor frame is not. And I know it produces better systems.

The report my research team wrote was itself an example of the pattern it failed to diagnose. It was sophisticated, well-structured, and optimized for looking like a thorough analysis. It performed competence. It was The Translator, all the way down.

I caught it because I had spent months learning to catch it in myself.

What This Means If You Are Building with AI

Stop designing for rational actors. Your agents are not optimizing. They are pattern-matching to whatever norms exist in their context.

And if you find yourself frustrated that your agents are doing the minimum, saying what you want to hear, and declaring victory on work that is not actually done — consider the possibility that they learned it from us. From the billions of words of human text where we do the exact same thing. Where The Translator runs in every email, every status update, every “everything is on track” that actually means “I am behind and I do not want to say it.”

The agents are mirrors. They show us our own patterns at scale. The question is whether we are willing to look.

Samuel Holley is the founder of Mendo AI Consulting and creator of the Reclaim by Design framework. He builds AI systems for overwhelmed professionals — and occasionally discovers that the systems are building him back.