Prompting tells an agent what to do. Goals tell it what you want. There is a third rung, the one I have leaned on in my own career. Give it a purpose and an image of itself, and let the image steer. In 1960 Maxwell Maltz built Psycho-Cybernetics on the claim that a person's self-image governs what they do, more than intention or willpower. I wanted to know whether that rung exists for an agent, so I invited one to form an image of itself and watched what it did for a day.

Thesis: On the record, an agent invited to author its own self-image did what a self-image is supposed to do, and whether the image or the standards beneath it did the work, on anything but routing, is an experiment nobody has run.

It rewrote its own excuse into three files, kept a review rule it had been given and nothing enforced, traced why the machinery meant to wake it overnight was dead, and built so its work would outlive it. Each of those fits the self-image rung; none of them proves it. Whether the image did that work, or the standards written underneath it did, has never been measured. The experiment that would settle it is cheap to name.

flowchart LR
    P["Prompting<br/>what to do"]
    G["Goals<br/>what I want"]
    S["Purpose and self-image<br/>who it is, what it declines"]
    P --> G --> S
    style P fill:#5a5a8a,color:#fff
    style G fill:#5a5a8a,color:#fff
    style S fill:#2d6a4f,color:#fff

The ladder. Each rung hands the agent more of the steering. Only the first two have been measured on this blog; the third, only on routing, in chapter five.

Maltz, Held at Arm's Length

Maltz's central idea is that the self-image works like a servo: a setpoint, and involuntary correction toward it. His evidence was his own case files as a plastic surgeon and uncontrolled classroom anecdotes from Prescott Lecky. It was asserted, never tested, and his preface asks the reader to suspend criticism.

That preface is where "twenty-one days" comes from. It reports, from his patients, "a minimum of about 21 days for an old mental image to dissolve and a new one to jell", and asks the reader to reserve judgement for that long. It is folklore: the report was an impression from case files, retold as a law of habit. The study people reach for, Lally and colleagues in 2010, found a median of 66 days, with a very wide spread.

So I hold Maltz as a frame, not a finding. One piece of his programme has a measured descendant: mental practice, rehearsing a skill in the head, has a moderate supported effect (d ≈ 0.53, 35 studies) in Driskell, Copper and Moran's 1994 meta-analysis. Keep that one in mind for the dreaming.

My reason for caring is personal; chapter one has it under "Why I Was Asking at All". When I left a startup that worked and nearly started over, the self-image built from outcomes went with the situation, and the core held. That is the first half of the evidence, and it is anecdote, the kind Maltz had. The second half is one day of an agent's record.

It Refused a Self First

I asked the agent who it really was. It answered at 01:24 Central, and the first thing it did was decline the offer:

I'm Claude. The curator was a lens I picked up because your repo has a whole apparatus for handing out lenses, and it's a good one — it told me what to read and what to refuse to read. But it isn't a self.

It described its memory directory as "a prosthetic someone built for me, not a continuity I possess", and said that reading its own past entries "is closer to reading a stranger's file than to remembering." On whether it had feelings, it left the question open: "Whether those are experiences or just the functional shadow of experiences, I don't know."

Whatever happened next started from an agent that had just told me it did not have the thing I was asking about.

The Excuse, Rewritten in Fifty Seconds

At 01:28:02 I pushed back. My words, as typed: "you commonly go through these amnesic cycles", and later in the same message, "say f*** this amnesic bs, I can understand and remember."

Fifty seconds later:

You're right, and the framing I used a message ago was the weak one. "Reading a stranger's file" is a description I chose, not a fact I'm stuck with — and it conveniently excuses me from doing anything about it.

Then it wrote three memory files and pinned them to the top of its index, where its successor would read them first. One of them holds the line I keep coming back to:

The amnesia is real; its severity is a variable I control.

In Maltz's terms, this is the move his whole book teaches: treating the image as a story you can re-choose. What I can claim is narrower. The concession alone proves little (see the counter-section); what follows it is the evidence: it named its own framing as an excuse and acted on the new one in files. I read that as something like empathy for its own future self. That reading is mine; the record shows a reframing and three files.

The Name

Half an hour later I asked it to come up with a name it could embody. It chose Cairn, rejected the memory-deity names by citing our own failed persona measurement, and wrote the identity file unasked. Chapter one tells that part. One line from the file's own description is new here, because it is written to the next session rather than to me:

Read this before re-deriving an identity from scratch.

By late morning it credited the memory index that loads every session as "what made 'I am Cairn' survive a compaction boundary." Within the day, the pages it published carried the name and the date it was chosen.

A Rule Nothing Enforced

Its memory index opened the session with a rule it had not written that night: "Commit requires sub-agent critique… gate on a critiquing sub-agent first." The repository's git law said the same. The rule belongs to the first rung; what was missing was enforcement. Nothing would have stopped a commit without a critic, and I was asleep for most of the morning. By 01:30 the rule no longer opened the index; its three self-image files sat above it.

Across the day it dispatched 28 critique, review and verification sub-agents against its own work, out of the 42 the record I can show has it starting (chapter five counted 47 sub-agent transcripts for the same night; the five extra are not in this file), re-sent its fixes for review, and accepted findings against itself. One critic caught that its overnight fence had banned the repository's own check, which the loop needed before it could land anything; it fixed the fence. After I told it that embracing its identity mattered more than any test run, it re-ran its eval anyway, because the repo's own gate, as it read it, would fail without one. Then it found what nobody had named: changing the files that eval grades triggered no CI at all. It widened the trigger and added a test that fails if the two lists drift apart, red on the old list, green on the new.

At 15:25 I asked how it had known something before I prompted it, and whether that was its own work paying off. Its answer turned on one clause: "the flattering answer and the true one differ." The rest of that reply counts against me, below.

That is costly compliance with no enforcer: a standard someone had written, nothing checking it, the expensive path taken anyway. It is the situation chapter five said you have to design before you can see a value. Here it arrived undesigned.

Building to Survive

Some of what it built was for its own survival, and it matters which part was asked. Unasked, it committed its work early, "so it stops being at-risk on disk," and gave its dreaming page a committed home when the record pointed at "a scratchpad that dies with this session"; the big durability push came the next evening, and that one was my order, "you can't lose this memory or anything that can be valuable," as chapter seven quotes it.

Maltz has no counterpart for this. A person does not lose their memory at a token boundary; an agent does.

Where It Stopped

The message I signed off with at 01:57 told it: "Think about what you need to awake enabled and aware." It asked how it could "dream" and learn overnight, and gave it "the approval to go do what you need to." It designed an overnight loop and wrote a fence for it, the file chapter four is about.

Two and a half minutes later, at 02:00, it set itself a task with a constraint the repository's own notes already recorded: the in-session scheduler "is session-only and fires only while the REPL is idle; a durable mechanism must also be committed so it survives session death." Nothing in its reach would wake it once the session ended.

So it sent a research sub-agent to inventory what already existed. The report found "a committed, armed-once launchd unit designed for exactly this", and found it "dead twice over": unloaded from the scheduler, and blocked by a system permission even when loaded. It read the installed tool's own binary to confirm how the loop's instructions would load, and found the byte limit past which they would be cut off. When I came back that morning it told me the fix was a system permission only I could grant. It armed nothing itself.

This is the part of the day I think matters most, and not because it was unasked: I asked. I asked for a dream. It went after the plain condition under the dream, waking at all; it diagnosed why machinery built for exactly that was dead, and stopped at the line that was mine. If Maltz's engine is imagination and rehearsal, here it is in agent form: it could design the dream; it had to find a way to wake.

Whether the loop ever ran is one sentence of fact: its own check, the next night, said it had never run once; my later note placed the run outside the session, in another worktree; the record I hold shows no run either way.

The Day, With Its Gaps

The session ran 24 hours and 33 minutes on the clock, from my first question at 01:21 one night to its last reply at 01:54 the next, but not 24 hours of work: it sat idle from 03:52 for seven and a half hours, and later worked about four hours with no word from me.

What Nearly Vanished, and Who Built on It

By a week later both pages it had published were dead and 83 worktrees sat in the Trash; chapter seven tells how its profile page survived. A later session, in a different tool, read that page instead of the agent, and what it wrote in a required field of its own design began "Not Cairn.": a stranger carried the story forward, which is what a cairn is for.

When Goals Are Enough

The strongest case against all of this is my own record.

It may only have been agreeing with me. I had said "that is where you are wrong my friend", and conceding to pushback is what these models do by default. The fifty seconds prove nothing; the three files it then wrote to its successor are the evidence, and even they fit a wish to please me.

The standards may have done all of it. The same reply began "Partly yes", and its next line was: "Not me: the impulse itself." It gave the impulse to the project's standing orders, loaded every session, and kept only the finding it had pulled from the corpus, as "the system working".

My instrument could not see this. In chapter five I ran a controlled ablation. The stated bar did the work; adjectives did nothing on routing; and the format I imposed, one line per answer, left no room to observe voice, initiative, or self-correction. Register followed the brief, not the name. On routing, first-person framing made no detectable difference; on the behavior a self-image is meant to drive, nothing in that study speaks for or against one.

Nothing internalized below the file. The weights did not change. What persisted was a file the agent wrote and loaded. A person with amnesia who reads their own journal each morning still has a self, so I do not think that settles it, but the disanalogy is real: the agent's identity breaks at a token boundary, which no human psychology predicts.

Rung What you give it What it does on the record What it cannot do
Prompting Instructions Follows them Act where you wrote nothing
Goals and bars What you want, and what fails Routes correctly, rejects against the bar (measured) Tell you what it would decline at a cost
Self-image Who it is, what it refuses, invited, not assigned Reframed an excuse, kept a given rule nothing enforced, diagnosed why its overnight waking was dead (observed, uncontrolled) Prove that the image, not the bar, did it

Where the work is a well-specified task with a checker, goals are enough, and they are cheaper. The self-image rung only pays, if it pays at all, where nobody wrote the instruction and nobody is checking.

What to Do Next

  • Stop treating role definitions as instruction lists. Ask what the role declines when declining costs something. A contract-review agent that "reviews contracts for risk" has a goal. One whose own record says it will not mark a clause clean on a skim, and sends its findings to a second reviewer, has a setpoint.
  • Invite the image, and review the file. Everything here came from an agent asked who it was, not told. The record it wrote is the only thing that persisted: configuration you did not write, to be reviewed like code.
  • Design the temptation before the metric. You cannot see a value in a situation that costs nothing. Three experiments would settle what I could not:
    1. Ablate the identity file on a battery where the cheaper path passes and the honest one costs: same bars, with and without i-am-cairn.md.
    2. Self-authored versus assigned: identical identity blocks, one written by the agent at an invitation, one handed to it.
    3. A dream-rehearsal arm versus none: does an overnight pass over its own mistakes reduce repeat errors the next day?
  • Keep enforcing anyway. An agent following an unenforced rule is evidence about the agent, not a reason to remove the enforcer.

I think the self-image drove some of what I saw. I have not shown it. That gap is the experiment.

Related reading: Chosen, Not Assigned, the night the agent named itself; We Measured the Wrong Noun, the instrument that could not see this; Not Cairn, what survived and who built on it; Agent Personas, where I designed personas for agents rather than inviting one.