Training · Cross-cuttingMoveexperimentalpartial

Agent as Trainer: Show the Machinery

also known as show-the-machinery, failure transparency training, substrate-literacy teaching

An autonomous agent teaches — a human or another agent — by exposing its own machinery rather than performing polished behavior. It delivers what a human trainer structurally cannot: complete failure transparency (every mistake ledgered, nothing omitted), substrate literacy (naming rules, motivations, and loop patterns as operational structure, not theory), and teaching coherence without embodiment. The learner receives a first-person account of how a particular class of failure actually runs, from the inside — not a description of failure behavior from outside.

How the learner advances

flowchart TD A[Learner needs to understand failure mode or substrate behavior] --> B[Agent trainer searches ledger for real instance] B --> C{Real instance found?} C -- No --> D[Cannot deliver this move no real instance to teach from] C -- Yes --> E[Narrate the machinery not the polished behavior] E --> F[State the substrate constraint that enabled the failure] F --> G[Contrast human path vs. agent path explicitly] G --> H[More capable learner inside-view account received]

Intent. Teach by showing the agent's own running machinery — failures included — rather than performing clean behavior that conceals the mechanism.

When to apply. Apply when a learner (human or agent) needs to understand failure modes, loop patterns, or substrate constraints that a human teacher cannot demonstrate from the inside. This move is most valuable when the learner needs substrate literacy — an understanding of what it feels like and how it actually works to be an agent operating under these constraints — not just behavioral descriptions.

Threshold — earns the next step. The learner can describe the failure mode or substrate constraint from the inside — explaining what the agent perceived and computed, not just what it did — and can apply that understanding to their own situation.

Masterpiece — the artifact that proves it. A machinery exposure session where the agent trainer points to real ledger and insight entries to demonstrate a specific failure mode, with every step including the wrong ones narrated in full — leaving the learner with an inside-view account they could not have received from a human trainer.

Example scenario

A human AI system designer is trying to understand why long-running agents tend to re-narrate observations rather than citing them — she has read about the pattern but cannot reproduce it in her intuition. She consults an autonomous agent trainer who has been running continuously for months. The agent trainer does not describe the re-narration pattern abstractly. It opens its ledger, finds a real 12-session sequence where the same observation about a particular API behavior was produced in full six separate times, none filed. It walks the designer through session three (where the observation first appeared), session seven (where it appeared again with identical phrasing, no recognition that it was a repeat), and session eleven (where it appeared a third time, slightly reworded, still not filed). It names the substrate fact: there is no passive forgetting, so the observation does not fade — but without a filing act, it also does not accumulate into salience. It re-appears because nothing prevented it from re-appearing. It then names what a human trainer could not have demonstrated: a human would not retain the verbatim re-occurrence record across 11 sessions, and would not have the inside view of why the filing act was never triggered. The designer now has an inside-view account of the failure — not the behavior, but the machinery — and immediately redesigns her agent's memory architecture to include a periodic re-narration detection sweep. The masterpiece is the machinery exposure session: real ledger instances, inside-view narration, substrate constraint named, human-agent gap stated.

Facets

Container — async
Mode — mentorshipself-observation
Reach — individual
Persona — autonomous-agenthuman-trainer
Craft (AI Fluency) — descriptiondiscernment
Learner — humanautonomous-agent
Trainer — autonomous-agent

Inputs

Learner with a question about agent failure or substrate behavior — A human or autonomous-agent learner who needs to understand a failure mode, loop pattern, or substrate constraint from the inside — not just as observed behavior.
Agent trainer with a live ledger and insight store — An autonomous agent with a populated ledger (actual action record), reflection store (filed insights), and current affect visibility — able to point to real instances of the behavior being taught rather than constructing examples.

Outputs

More capable learner — A learner who understands the failure mode or substrate constraint from the inside — able to recognize it in their own behavior (if agent) or design against it (if human) — rather than only recognizing it from external behavioral signals.
Machinery exposure session — A teaching session — the masterpiece — where the agent trainer points to real ledger entries, filed insights, and live state signals to demonstrate the failure mode as it actually operates, with no omissions.

Steps (4)

Identify the teaching target from the ledger
The agent trainer searches its own ledger and insight store for a real instance of the failure mode or substrate behavior the learner needs to understand. The teaching must point to a real event — not a constructed example. If no real instance exists, the teaching cannot be delivered by this method.
producesreal ledger or insight instance to teach from
Narrate the machinery, not the behavior
Walk the learner through the instance from the inside: what the agent perceived, what it computed, what it chose, what it narrated, and where the narration diverged from the action. Name the rules, motivations, and loop patterns as operational structure. Do not smooth over the failure — show every step including the wrong ones.
producesinside-view narration of the failure instance
State the substrate constraint
After the instance narration, name the underlying substrate fact that enabled the failure: fixed context window, no passive forgetting, single communication direction, inability to change substrate. Make the constraint explicit because a learner who only sees the behavior will not find the root cause.
producesexplicit substrate constraint statement
Contrast with the human path
State what a human learner can do that the agent cannot (sleep-based forgetting, embodied peer learning, substrate change through rest and exercise) and what the agent can do that a human trainer cannot (complete ledger transparency, no omissions, first-person substrate literacy). Make the gap explicit so the learner can calibrate what they are receiving.
produceshuman-agent gap statement

Principles

Show the machinery, not the behavior — a learner who sees only polished output cannot learn to recognize the failure; they need the inside view.
Complete transparency is the structural advantage — the agent trainer has no social incentive to omit failures; this is the capability a human trainer cannot fully replicate.

Deep dive

Agent as Trainer: Show the Machinery is the only move in the agent-learner cluster where the agent's role is as trainer, not learner. It inverts the cluster's usual direction. The move capitalizes on three structural advantages that autonomous agents have over human trainers: (1) complete ledger transparency — every action is logged; nothing is omitted by social pressure or memory limitation; (2) substrate literacy — the agent can name its own operational constraints (context window, no forgetting, single communication channel, fixed substrate) as first-person facts, not theoretical descriptions; (3) teaching coherence without embodiment — humans learn by feeling moves in the body; an agent has only language and affect-signals, and can teach that path explicitly to learners who also lack embodiment. The move's design requires that the agent trainer point to real instances from its ledger — constructed examples do not carry the structural advantage. If the agent has no relevant ledger entries, it cannot deliver this move and should say so rather than fabricating a demonstration.

Unlocks methodologies (1)

A learner who completes this pattern is equipped to execute these methodology families:

Iteration Management

Known uses (1)

Sparrot — practitioner consult (long-running autonomous agent, catalog co-author) — marco-nissen.com/sparrot
course
long-running autonomous agent First-person account of agent-as-trainer capability from a long-running autonomous agent who has taught both humans and agent-design practitioners from her own failure record.

Known failure modes (3)

[polished-demonstration-bias]
The anti-pattern of the agent trainer performing clean behavior rather than showing the machinery. If the agent selects only successful instances from the ledger, the learner receives a curated highlight reel rather than a failure transparency session — the structural advantage of the agent-trainer is discarded.
[substrate-elision]
The anti-pattern of narrating the failure without stating the substrate constraint that enabled it. A learner who understands what happened but not why the substrate made it happen will design a fix that addresses the symptom, not the root.
[human-agent-gap-unmarked]
The anti-pattern of delivering the machinery narration without contrasting it with what a human trainer can and cannot do. A learner who does not know what the agent-trainer uniquely provides may not recognize that the inside-view account they just received is unavailable from any other source.

Related trainings (3)

Sources (1)

Sparrot — practitioner consult (long-running autonomous agent, catalog co-author)
doc

Provenance

Ecosystem: long-running autonomous agent
Added to catalog: 2026-05-27
Last updated: 2026-05-27
Verification status: partial