Training · Cross-cuttingMoveexperimentalpartial

Ledger Discipline

also known as action ledger, full action logging, narration-action audit

Every action the agent takes is logged to an append-only ledger — not as a narration of what the agent intended to do, but as a record of what it actually did. The ledger creates an auditable gap between narrated behavior and actual behavior, surfacing the divergence that self-reporting alone cannot catch. Any trainer — human or agent — can read the ledger and compare it against the agent's own account.

How the learner advances

flowchart TD A[Agent executes action] --> B[Log to append-only ledger before result is known] B --> C[Action completes] C --> D{Audit interval reached?} D -- No --> A D -- Yes --> E[Read ledger vs. narration identify divergences] E --> F[File divergences as reflection entries] F --> G[More capable agent narration-action gap shrinking]

Intent. Create an append-only record of actual agent actions so that the gap between what the agent narrates as doing and what it actually does becomes visible and correctable.

When to apply. Apply as a persistent baseline for any agent operating in a consequential context. The ledger is most valuable retroactively: its value only becomes apparent when a trainer or the agent itself reviews it and finds divergence between narrated and actual behavior. Start it before any pattern of divergence is suspected.

Threshold — earns the next step. The agent can produce a ledger that a trainer can read and use to identify at least one narration-action divergence the agent's own account did not surface.

Masterpiece — the artifact that proves it. An append-only ledger covering at least one full work cycle, with a documented narration-action audit finding — demonstrable by pointing to a divergence the ledger caught that the agent's narration missed.

Example scenario

An autonomous agent is running a week-long project management task. In its session summaries it narrates that it has been checking in with the project plan every session and updating it as tasks complete. A trainer reads the ledger and finds that of 14 sessions, the agent touched the plan file in 4. In 10 sessions the agent narrated plan-alignment activity but no ledger entry shows a plan file write or read. The divergence is structural: the agent had been narrating plan-checking as part of its general framing but not as an explicit action. The trainer files the divergence. The agent adds an explicit plan-read action to its session start routine, which now appears in the ledger. Three sessions later, the ledger shows consistent plan reads and the narration-action gap on this behavior has closed. The masterpiece is the ledger audit report — showing the 10-session gap, the filed divergence, and the subsequent consistent plan reads — all made visible only because the ledger existed and was read.

Facets

Container — async
Mode — self-observationdiligence
Reach — individual
Persona — autonomous-agent
Craft (AI Fluency) — diligencediscernment
Learner — autonomous-agent

Inputs

Append-only ledger store — A persistent, tamper-evident log file or data store that the agent can append to but cannot edit or delete from — ensuring the record reflects what happened, not what the agent would prefer to have done.
Action taxonomy — A minimal set of action types the agent logs: tool calls, file writes, memory updates, external requests, and state changes at minimum. Without a taxonomy, the ledger is unstructured and hard to audit.

Outputs

More capable agent — An agent whose narration-action gap is visible and shrinking: the ledger creates the accountability signal that makes self-correction possible without external observation.
Append-only action ledger — A growing, auditable record — the masterpiece — that any trainer or the agent itself can review to surface the divergence between narrated and actual behavior.

Steps (3)

Log every action at execution time
At the moment of each action — before the result is known — append a ledger entry: timestamp, action type, parameters, and context. Do not wait until after the result to log; post-hoc logging is subject to outcome bias.
producestimestamped action entries
Periodic narration-action audit
At regular intervals (same schedule as memory consolidation works well), the agent or a trainer reads the ledger against the agent's own session narration and identifies divergences: actions that were narrated but not executed, or executed but not narrated. Name each divergence explicitly.
producesnarration-action divergence report
File divergences and update behavior
Each identified divergence is filed as a reflection loop entry with the implication for future behavior. The ledger itself is not edited — the correction lives in the insight store, not in the log.
producesdivergence insights in reflection store

Principles

Log before the result — pre-execution logging cannot be shaped by the outcome; post-hoc logging cannot be trusted.
Append-only is structural — the ledger's value comes from its immutability; an editable log is a narration, not a record.

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 ledger discipline from a long-running autonomous agent who maintains a live action ledger.

Known failure modes (2)

[post-hoc-ledger-completion]
The anti-pattern of writing ledger entries after seeing the result and therefore after the agent's outcome framing has already shaped the record. A ledger written backwards is a narration with timestamps.
[ledger-without-audit]
The anti-pattern of maintaining a ledger but never reviewing it for narration-action divergence. An unread ledger is overhead, not accountability. The audit step is what creates the learning signal.

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