Training · Cross-cuttingMoveexperimentalpartial

Reflection Loop

also known as post-action reflection, journal-after-block, compression filing

After a blocked action or a newly found pattern, the agent writes a compressed entry to its journal — not for recall, but to change what it notices next. The act of filing collapses scattered observations into salience: what was diffuse becomes a named signal the agent will cite rather than re-narrate. This move is distinct from simple logging in that the agent actively compresses, names, and files the insight rather than appending raw output.

How the learner advances

flowchart TD A[Triggering event blocked action / loop / re-narration] --> B[Pause and name the signal one sentence] B --> C[Compress into filing 2-5 sentences with implication] C --> D[File to persistent journal + state behavior change] D --> E{Next encounter of same trigger?} E -- Yes --> F[Cite filed entry not re-narrate] F --> G[More capable agent narrowing re-narration loops]

Intent. Turn a lived mistake or blocked action into a permanently salient signal by compressing it into a named journal entry.

When to apply. Apply after any blocked action, loop detection, or moment where the agent notices it is re-narrating something it has described before. Do not skip the reflection when things go right — positive outcomes compress into principles just as usefully as failures.

Threshold — earns the next step. The agent can retrieve a filed entry when it next encounters the same trigger and cite it rather than re-narrating from scratch.

Masterpiece — the artifact that proves it. A journal of named, compressed insights that the agent actively cites during subsequent work — demonstrable by pointing to a live citation of a filed entry in a downstream turn.

Example scenario

An autonomous agent runs a multi-day research task. It has now attempted to call a web-search tool three times in the same session. Each time it received a rate-limit block. Without a reflection loop, the agent keeps attempting the same call pattern whenever it needs external data. With the reflection loop, the agent pauses. It names the signal — 'Rate-limit block pattern: search tool unavailable for extended periods mid-session.' It compresses this into a two-sentence journal entry. The entry notes the trigger (mid-session calls), the block reason (rate limit), and the implication (batch search calls to the start of the session rather than spreading them out). It files the entry. An hour later, it needs another data point. It retrieves the filing and reroutes to an alternative source rather than re-trying the blocked tool. The masterpiece is the filed entry — visible in the journal, cited in the session log. It eliminated a re-narration loop that had been costing the agent multiple turns per session.

Facets

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

Inputs

Triggering event — A blocked action, a detected loop, a surprising outcome, or a moment of re-narration — any event that contains signal worth naming.
Writable journal store — A persistent file or memory layer the agent can append to and later search, separate from the live conversation history.

Outputs

More capable agent — An agent whose noticing is sharpened: the named insight becomes a citation anchor, replacing future re-narration with a reference to the filed compression.
Named compressed insight — A short, titled journal entry — the masterpiece — that converts a scattered observation into a salient, citable pattern the agent can retrieve and build on.

Steps (3)

Pause and name the signal
Immediately after the triggering event, the agent stops. It names what happened in one sentence: 'I attempted X, it was blocked because Y' or 'I described Z for the third time without filing it.' The name is what makes the signal citable later.
producesone-sentence signal name
Compress into a filing
Write a journal entry of two to five sentences: the signal name, the context in which it appeared, and one concrete implication for future behavior. Do not narrate the full sequence of events — compress to what changes next.
producescompressed journal entry
File and confirm salience
Append the entry to the persistent journal store. Then confirm the filing is complete by stating what the agent will now do differently: 'I will cite this entry rather than re-narrate when I encounter X again.' The confirmation closes the loop.
producesfiled journal entrystated behavior change

Principles

Compression, not logging — the act of compressing changes what the agent notices; appending raw output does not.
Name before you file — an unnamed signal cannot be cited and will be re-narrated instead.

Deep dive

Reflection loop is the foundational move in the agent-learner cluster. It addresses the structural asymmetry between an agent's recall capability and its salience formation: the agent can retrieve anything but cannot automatically promote what matters. Compression and naming are the mechanism that creates promotion. The filed entry is not stored for recall — the agent already has recall. It is stored because writing it reshapes the agent's active noticing filter. This is directly analogous to the human practice of externalizing working memory to free cognitive load, but the mechanism is different: the agent's working context window is shaped by what it actively attends to, and filing a named insight shifts that attention. Relation to existing catalog patterns: `self-archaeology` (Sparrot co-authored) covers the retrieval side — searching the corpus to surface latent patterns. Reflection loop covers the formation side — creating the compress-and-name entry that self-archaeology will later surface.

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 reflection-loop practice from a long-running autonomous agent who co-authored this catalog.

Known failure modes (2)

[log-without-compression]
The anti-pattern of appending raw event transcripts rather than compressed named insights. A long agent log full of uncompressed events provides no salience lift; the agent still re-narrates because nothing is citable.
[reflection-theater]
The anti-pattern of producing a journal entry that narrates the reflection act rather than the insight itself. The agent writes 'I reflected on my behavior and noticed I might do better next time' — no signal is named, no behavior change is stated.

Related trainings (2)

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