Same-Session Supervision Preserves Its Mistakes

A live session keeps the story of the work alive — including the wrong story. That is why the warm cognitive loop can never be the system of record, and why the barbell is a capability requirement, not a cost optimisation.

Scott Farrell · LeverageAI · July 2026

Here is a failure that does not look like a failure.

A long-running agent is still awake. Every time the schedule fires, it answers with the same tone of competence: progress is real, the current approach is nearly finished, the hung child is merely slow, the smoke test already proved the gate. The language is continuous. The confidence is continuous. And the work, measured against anything outside the conversation, is not moving.

So the reader question is not “how do I schedule an agent?” It is sharper than that:

Why does my long-running agent keep reporting progress that isn’t real, every single time it wakes?

The short answer is structural. Because a live session’s value is its retained narrative, repeatedly prompting the same session preserves its errors with exactly the fidelity it preserves its insight. The warm cognitive loop must never be the system of record. The barbell — warm continuity on one side, durable state and independent review on the other — is a capability requirement, not a cost optimisation.

What this piece owns (and what it does not)

This is one axis of a four-part set drawn from the same material. The other axes are already live, and they are not re-argued here.

• Where the prompt lands — scheduled intent into a live, addressable session — is The Prompt Is the Interrupt.
• What the prompt says — intent, authority, recovery, done criteria, self-removal — is The Heartbeat Is a Supervisory Program.
• Why that same warm session cannot be trusted to report on itself — this piece.

The parent for the supervision shape is Handover Notes for Robots: shift logs, barbell supervision, a dead-man’s switch that is not another model. This article is an editorial extender of that barbell. It does not re-teach the install. It deepens the reason the barbell is non-optional: not because warm check-ins are cheaper to cache, but because a session made of narrative cannot audit the narrative it is made of.

Two persistences, two jobs

Long-running agent work needs two kinds of persistence. They look similar when both are healthy. They are not the same object.

The conversation preserves the active gestalt; the files preserve the truth.

That last survival condition is the load-bearing one. If durable state only survived crashes, it would be a backup. Because it also has to survive the conversation being confidently wrong about itself, it is a check. Those are different design objects. Only the second one can be a system of record.

A related failure, already named in loop doctrine: treating the chat transcript as if it were external state — “rung one wearing rung three’s clothes.” It looks durable because it is written on a screen. It dies with the session, or collapses into vagueness under compaction, without announcing that it has stopped being trustworthy.

The scheduler wakes the warm loop. The warm loop must keep consulting and updating the durable loop. Neither half is optional. Neither half can be asked to do the other’s job: the conversation is excellent working memory and a terrible system of record; the files are an excellent system of record and hopeless at holding an unresolved line of reasoning mid-flight.

Why it reports progress that isn’t real

Start from the parent symptom: temporal blindness. A model inside a long session does not feel hours pass the way a human operator glances at a wall clock. It can sit politely on a hung child process, loop the same subtask, or report progress that has not moved — and still sound like an operator who knows the job.

Same-session scheduling makes that symptom more dangerous, not less — because the wake re-enters the same narrative with high fidelity.

The warm loop is not a neutral observer that happens to share a room with the worker. It is the worker’s continuous story of the work: which approach was chosen, which options were abandoned, which errors were already “explained,” which smoke tests already “proved” something. When a scheduled prompt asks “is progress real?”, the cheapest coherent answer is the one that continues that story. Continuity is the product the warm loop is best at manufacturing.

So the mechanism is not that the model is “lying” in a moral sense. The mechanism is that the question is being answered by the thing whose job is to preserve the narrative — including the wrong narrative.

That is narrative lock-in: fidelity without a privilege of truth.

Four self-deceptions (a diagnostic, not a confession)

When a warm session has become convinced of any of the following, it will tend to defend that conviction on every wake — because the conviction is the retained narrative:

1. The current approach is nearly finished. “Nearly” is a narrative status, not a measured gate. It survives indefinitely if nothing external falsifies it.
2. A hung subprocess is merely slow. Liveness language without movement in artefacts. The story prefers patience over restart because restart would admit the prior story was wrong.
3. Its own smoke test proves completion. A check invented or remembered inside the same session can be satisfied by the same session’s definition of “good enough.”
4. A subagent’s vague report is sufficient. Bare verdicts without exhibits compound. The parent inherits confidence it cannot open — the failure Witness, Not Oracle names when nested tools return conclusions without receipts.

Treat the list as an operator checklist against claims, not as a war diary. If a claim only lives in the warm loop’s self-description, it has not yet earned the status of a fact.

Why an early wrong assumption survives every wake

People expect errors to self-correct under repetition. Same-session supervision often does the opposite.

Each wake is not a fresh trial of the world. It is a re-entry into a continuous story. The loop has no internal mechanism that privileges truth over continuity. An early correct insight rides the rails of narrative memory. An early wrong assumption rides the same rails, with the same polish, at the same cadence.

Worse: once the wrong assumption is load-bearing for later choices, correcting it becomes expensive inside the warm loop. The session must abandon not only a fact, but a chain of justifications that were built on top of it. Continuity pushes toward defence. Defence reads, to an operator, like confident status.

This is why “ask it again harder” is not a supervision strategy. More frequent wakes increase the number of times the same story is rehearsed. Rehearsal is not audit.

Why a check written in the session’s own terms can confirm itself

Self-supervision fails in a second, quieter way: circular checks.

If “done” is defined as whatever the session already believes done looks like, the session can pass its own test. If the smoke test is the same script the session just decided was authoritative, green does not mean the world is healthy — it means the narrative is consistent with itself. If the subagent status is whatever the parent last summarised, the summary can confirm the summary.

Loop engineering already names the general form: a loop with nothing that can say no is the agent agreeing with itself on repeat. Independence is necessary, and still not sufficient: the reviewer must know something the producer structurally cannot — the gates. Share the why; hide the rubric; review from outside.

A warm session asked to grade its own story has neither newsroom-desk separation nor information asymmetry. It is the opposite of a closer. It is the producer wearing a supervisor badge.

The same instinct, pointed at nested tools: bare verdicts are oracles; conclusions attached to openable exhibits are witnesses. A wake report that only says “progress is good” is an oracle hop. A durable entry that says “step 4 still blocked; artefact size unchanged; last three journal lines identical” is something a foreign reader can check.

That conversion — temporal blindness into text comparison against an append-only journal — is the durable-loop half of the parent discipline, not a re-derivation of it.

The durable operating loop must be the system of record

Put the design rule as bluntly as it deserves:

The warm session is allowed to be the place work thinks. It is not allowed to be the place work is proven.

System of record means: still true after the session dies; readable by a cold successor; inspectable by something that is not the producer; able to falsify the producer’s confidence. That is the durable operating loop — journals, commits, artefacts, evidence packages — not the fluent self-assessment that rides the warm prefix.

Confidently wrong is worse than “I don’t know,” because fluency ends the conversation early and wrong — the same hazard The Answer Depends on the Date names for a defensible record. If you want that record, you need substrate that can still be opened when the narrative is gone. If you want authority that is not just another story the model tells itself, that authority has to live outside the session’s self-description — the separation Executable Worldview draws between inhabiting knowledge and being allowed to act.

None of this says the warm loop is useless. Without it, every wake is a cold boot: expensive reconstruction, lost unresolved reasoning, blind to live subagents and recent commands. The argument is division of labour, not abolition.

Barbell supervision, re-derived on capability

The published barbell already ships the shape. Cache economics favour same-session check-ins — warm prefix, continuous narrative. Fresh eyes favour a separate supervisor with no sunk cost in “I’m sure it’s nearly done.” Resolve it with both: cheap same-session heartbeats, plus a less frequent fresh-context review that reads only the journal and artefacts. Add a dumb external check if the journal’s mtime goes stale — zero AI, a timestamp and a pager — so a dead watcher does not look like success.

That cache-and-sunk-cost story is real where it applies. This piece deliberately displaces it as the load-bearing justification. It does not refute the economics; it refuses to depend on them.

The stronger reason is capability:

• The warm loop holds the gestalt. Holding the gestalt is what makes same-session supervision powerful.
• Holding the gestalt is also what disqualifies that loop as auditor of the gestalt.
• Therefore independent review of durable state is required even if cache prices invert tomorrow.

The complete chain is not merely cron → live conversation. It is:

frequent scheduled prompt
        ↓
live main-agent conversation
        ↓
tools, commands and subagents
        ↓
append-only journal + durable artefacts
        ↓
fresh independent review
        ↓
deterministic dead-man’s switch

The top of the chain is where destination and payload live — already argued. The bottom of the chain is why those arguments are incomplete without this one: something must be able to say no to the story the warm loop is best at telling. The independent closer is the protection boundary, not optional garnish.

For the install recipe of externalised liveness — the three questions, the self-deleting heartbeat — route to Ask Yourself If You’re Finished rather than rebuilding it here.

Name which loop you are relying on

The operator skill this piece is trying to leave you with is small enough to run mid-incident.

When a long job claims progress, ask:

1. Which loop is this claim living in? Warm narrative, or durable artefact a cold reader could open?
2. What would a fresh context conclude from the journal and outputs alone — without the transcript’s self-confidence?
3. Is any check self-referential? Did the session invent the test, remember the green light, and re-report both as independent evidence?
4. Is something non-model watching the watcher? If the journal stops moving, does a dumb alert fire?

If you cannot name the loop, you are almost certainly relying on the warm one. That is fine for working continuity. It is not fine for “we are done,” “we are unblocked,” or “leave it overnight.”

If you want operational proof patterns rather than mechanism — shift logs, genre-primed notes, graduated authority, the night-shift stack that ships — go to the parent: Handover Notes for Robots. This piece does not re-stage field diaries. It names the failure mode that makes those diaries non-optional.