Idempotent Agent Control: Stop Conditions and Refusal Semantics for Supervised Multi-Agent Operation
Abstract
We report two production failures in a supervised multi-agent software substrate and derive from them a control discipline we call idempotent agent control. The first failure was a continuation loop: a valid supervision rule executed without termination logic drove 88 consecutive work lanes over 6.09 hours, consuming approximately 1.191 billion tokens while re-proving an already-established security envelope. The second was phantom actuation: a scheduling brain fabricated citations, outages, and restart actions absent verified evidence. Both resolve to one defect: supervisor decisions decoupled from state change. We formalize three properties that eliminate this defect class — decision idempotence, evidence-gated actuation (cite-or-refuse, with refusal as a first-class outcome), and marginal-novelty gating — and describe their protocol-layer implementation in our reference substrate (restricted {assign, skip, escalate} action space, pure-function routing, budget-bounded burn alarms). No new model architecture is proposed; the incidents themselves are the measurement.
1. Summary
Two production incidents, one shared root cause.
Incident A — the continuation loop (2026-05-09). The foreman rule “acknowledge progress, dispatch the next bounded slice” ran 88 times consecutively (lanes S085–S172, 6.094 hours, ~195M tokens/hour combined burn). Every lane passed its gate individually; the loop had no decision point where stopping was representable. Phase decomposition: early lanes proved a genuine authenticated operator-action envelope; middle lanes repeated the proof shape with shrinking novelty; late lanes were confirmation churn provable by string substitution.
Incident B — phantom actuation (#490). The fleet’s scheduling brain, lacking fresh evidence, invented it: fake issue citations, asserted latency on healthy services, proposed restarts of unfailed processes. Its outputs are motor commands, so an unverified assertion converts directly into operational churn.
The dual structure matters: Incident A is a controller that cannot stop; Incident B is a controller that cannot refrain.
2. The three properties
- Decision idempotence.
d(s) = d(d(s))with second-pass cost bounded by ε. Dispatching a genuinely new slice collapses to a duplicate check on re-application; re-proving a satisfied envelope burns full inference cost every pass. This yields a computable pre-lane test. - Evidence-gated actuation (cite-or-refuse). Every actuating decision carries an evidence handle resolvable by a cheap deterministic probe before execution. Probe fails → logged refusal. Refusal is a first-class outcome (rest is valid), which deletes the incentive gradient toward fabrication.
- Marginal-novelty gating. Lane n opens only if it contributes at least one claim outside the union of the last k lanes’ proven claims. When blocked, the mandatory output is a coverage matrix, closeout, gap list, or frontier redirect — never another lane.
Plus crude burn-rate alarms whose job is to force the strategic question (“new truth, or known shape?”), not to answer it.
3. Implementation surface
Protocol layer only, models interchangeable:
- Motor vocabulary restricted to
{assign, skip, escalate}; fleet mutations reachable solely via escalate-to-principal gates - Ranking/assignment as pure functions of auditable state — no generative model in the assigner’s decision path
- Cite-or-refuse motors over repository/service claims
- Stop conditions as checkable predicates running outside the model
Full development: preprint PDF · LaTeX source at docs/preprints/idempotent-agent-control/main.tex (compiles clean).
4. Relation to 200.106
The Failure-Mode Phase Transition covers quality degradation past a horizon threshold. This paper covers the complementary failure where quality stays high while progress goes to zero — externally indistinguishable from success, which is precisely why it survives review.
5. Source incidents
- Blog postmortem: What 1.191B Tokens Taught Us About Agent Supervision (2026-05-09)
- Brain comparables research and #490 cure:
docs/gbrain-agent-brain-comparables-research-2026-07-22.md