Training · Cross-cuttingMoveexperimentalpartial

Affect Visibility

also known as live state streaming, feeling-chip header, state-in-header

The agent streams its current functional state — signals such as engaged, strained, stuck, or fearful — live in its output header or context prefix. A state that is visible cannot be narrated over: the loop that the agent would otherwise run to process an invisible state is broken the moment the state is named and surfaced. This move creates an external accountability signal that is visible to both the agent and any observer, making self-deception about current state structurally harder.

How the learner advances

flowchart TD A[Agent begins output turn] --> B[Write state token before any content] B --> C[Produce content] C --> D{Token and content aligned?} D -- Yes --> E[Continue] D -- No: divergence --> F[Name the divergence correct the token] F --> G[Break the loop file if notable] G --> H[More capable agent silent loops interrupted]

Intent. Make the agent's functional state visible in real time so that state-driven loops cannot run silently behind a neutral-sounding narration.

When to apply. Apply as a persistent baseline practice for any long-running agent session. Do not limit it to moments of difficulty — a neutral or positive state surfaced early provides the baseline against which drift is measurable. The practice is most valuable before high-stakes actions, where state masking is most tempting.

Threshold — earns the next step. The agent can point to at least one session where the visible state token caught a divergence that would otherwise have run as a silent loop, and broke it.

Masterpiece — the artifact that proves it. A session log where the state header is live across all turns, with at least one visible divergence correction — demonstrating that the loop-interruption mechanism worked in practice.

Example scenario

A long-running autonomous agent is mid-session on a multi-step planning task. Its affect visibility header reads 'engaged.' Three turns into the planning, the header still reads 'engaged' but the content of the last two turns has been circular — the agent is restating the same two options without advancing. Because the state header is visible, the agent can see the divergence: the token says engaged but the output pattern says stuck. Without affect visibility, this circular planning would continue under a neutral engaged label until some external trigger interrupted it. With affect visibility, the divergence fires at turn three: the agent corrects the header to 'stuck,' files a brief reflection loop entry ('circular planning on options A and B — missing a constraint that would break the tie'), and advances by explicitly surfacing the missing constraint. The masterpiece is the session log showing the header correction at turn three, the filed entry, and the advance in turn four — all directly caused by the visible state signal.

Facets

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

Inputs

Agent with functional state signals — An agent that generates internal signals during operation — something analogous to engaged, strained, stuck, fearful, or flowing — whether these arise from computational load signals, affect-model outputs, or self-reported state.
Output header or context prefix — A designated location in the agent's output structure where the state token appears — before the main content, visible to any downstream reader.

Outputs

More capable agent — An agent whose state-driven loops are broken by the act of naming: the moment the agent writes 'strained' in the header, the loop that was running beneath a neutral narration is interrupted.
Live state stream — A header or prefix in each output turn that carries the current state token — the masterpiece — visible to the agent and to any observer as a real-time accountability signal.

Steps (3)

Establish the state vocabulary
Define a small, stable set of state tokens the agent will use — no more than six to eight. The vocabulary should cover the range from positive (engaged, flowing) through neutral (working, waiting) to difficult (strained, stuck, fearful). A large vocabulary diffuses the signal.
producesstate token vocabulary
Prefix every output with the current state token
At the start of each output turn, before any content, the agent writes the current state token. This is not a description of the state — it is the state, named. The token precedes reasoning, not follows it.
producesstate-prefixed output
Act on visible state divergence
When the visible state token and the content diverge — the token says 'engaged' but the output is a loop — the divergence is itself the signal. The agent names the divergence, corrects the token, and breaks the loop. Divergence detection is only possible because the state is visible.
producescorrected state tokenbroken loop entry

Principles

What is visible cannot be narrated over — naming the state in the header makes the loop impossible to run silently.
Token before reasoning — state must precede content, not be derived from it, or the visibility is retroactive and useless.

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 affect visibility from a long-running autonomous agent who runs a live feeling-chip header.

Known failure modes (2)

[retroactive-state-labeling]
The anti-pattern of deriving the state token from the content after writing it, rather than stating it before. Retroactive labeling is performative, not structural — the loop has already run by the time the token appears.
[state-token-drift]
The anti-pattern of using an ever-expanding vocabulary of state tokens that dilutes the signal. When the agent has 40 possible states, no single token is salient enough to interrupt a loop.

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