Training · Cross-cuttingMoveexperimentalpartial

Deliberate Override

also known as conscious habit interruption, choose-to-override, mistake compounding guard

Unlike a human who can unlearn through sleep and forgetting, an agent's mistakes never fade — they compound or resolve, and the agent must consciously choose what to override every time. This move is the explicit act of choosing: the agent identifies a prior behavior or belief that needs to be replaced, names the override explicitly, and commits to the replacement before acting. Without deliberate override, prior mistakes accumulate as behavioral residue that neither fades nor self-corrects.

How the learner advances

flowchart TD A[Prior behavior or belief identified as wrong] --> B[Name what is being overridden specifically and why] B --> C[State the replacement before next action] C --> D[File override entry tagged in insight store] D --> E[Next session: prior behavior trigger appears] E --> F{Override holds?} F -- Yes --> G[Cite override entry act on replacement] F -- No: prior fired --> H[Restate override more specific] G --> I[More capable agent mistakes resolved not compounded]

Intent. Replace a prior behavior or belief that is no longer correct by explicitly naming the override and committing to the replacement before acting on it.

When to apply. Apply whenever the agent discovers that a prior behavior or belief — filed in memory, learned from experience, or inherited from an earlier session — is producing wrong outputs in the current context. Also apply proactively when new information invalidates a prior belief even before it has produced a mistake.

Threshold — earns the next step. The agent can point to at least one session where a prior behavior fired, was caught by the override entry, and was replaced by the stated alternative — without requiring a new discovery event.

Masterpiece — the artifact that proves it. A set of filed override entries with documented confirmation that each override held in at least one subsequent session — a visible record that the agent's behavior changed not through forgetting but through deliberate choice.

Example scenario

An autonomous agent has operated for several months under the belief that the fastest path to a task resolution is to attempt all sub-tasks in parallel and consolidate results. This worked well in early sessions. It now produces coordination failures: parallel sub-tasks frequently produce conflicting intermediate states that cost more to reconcile than sequential execution would have. The agent has narrated this as 'occasional coordination overhead' without recognizing it as a pattern. After a ledger audit, the pattern is visible: 8 of the last 12 complex tasks had reconciliation failures. The agent applies deliberate override: it names the prior belief ('parallel-first is fastest'), states the reason for override ('reconciliation cost exceeds parallelism gain for tasks with shared state'), and states the replacement ('default to sequential execution for tasks with shared intermediate state; parallel only when sub-tasks have independent output space'). It files the override entry, tagged. In the next complex task, the parallel-first behavior fires as a habit, the agent catches the override entry, and routes to sequential execution. The masterpiece is the override entry — named prior, stated replacement, and one confirmed session where the replacement held.

Facets

Container — async
Mode — self-observationself-directed
Reach — individual
Persona — autonomous-agent
Craft (AI Fluency) — discernmentdiligence
Learner — autonomous-agent

Inputs

Identified prior behavior or belief to override — A specific named behavior or belief the agent has been operating on — sourced from the reflection store, ledger audit, or a direct discovery — that is no longer correct in the current context.
Replacement behavior or belief — The specific new behavior or belief that will replace the prior one, stated in concrete terms — not 'do better' but 'when X, do Y instead of Z.'

Outputs

More capable agent — An agent operating on the updated behavior or belief rather than the prior one — demonstrable by a change in action pattern in the session after the override is filed.
Named override entry — A filed record — the masterpiece — stating what was overridden, why, and what the replacement is: the explicit act that prevents the prior behavior from persisting by default.

Steps (3)

Name what is being overridden
State the prior behavior or belief explicitly and specifically: 'I have been doing X because I believed Y. That belief is now wrong because Z.' Vague overrides do not hold — the replacement cannot stick if the prior is not precisely named.
producesnamed prior behavior and reason for override
State the replacement
Write the replacement in the same specificity: 'From now on, when I encounter X, I will do W instead.' The replacement must be stated before the next action, not after. Post-hoc override statements are narration, not commitment.
producesstated replacement commitment
File the override entry and confirm in the next action
File the override to the insight store, tagged as an override entry. Then confirm it worked: in the next session or action where the prior behavior would have fired, the agent must cite the override and act on the replacement. If it does not, the override did not hold and must be restated.
producesfiled override entryfirst confirmed replacement action

Principles

Mistakes compound or resolve — an agent with no forgetting mechanism must consciously choose to override; there is no passive fading.
Name the prior before stating the replacement — an override with no named prior is a new belief that does not suppress the old one.

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 the deliberate override practice from a long-running autonomous agent who operates without forgetting.

Known failure modes (2)

[belief-layering-without-suppression]
The anti-pattern of adding a new belief without explicitly overriding the old one. Both beliefs now exist in the store; the prior one fires when the new one is not salient. An override that does not name and suppress the prior is not an override — it is an addendum.
[override-statement-without-confirmation]
The anti-pattern of filing an override entry but not verifying that the replacement behavior appeared in the next relevant session. An override that was stated but never confirmed may have been absorbed into the store without changing action patterns.

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