Agent Council

A function-specific layer of specialized review agents governing execution agent outputs before escalation to the human Steward — the agentic equivalent of the management function, handling quality review, error detection, misinterpretation resolution, and escalation triage between the Execution Layer and the human Judgment Layer.

The Agent Council refines the existing Execution Layer / Judgment Layer binary into a three-tier architecture. The existing binary assigns everything the Execution Layer cannot handle to the human Judgment Layer — but not all judgment requires human judgment. The Agent Council governs T2: the quality review, consistency checking, escalation triage, and pattern-learning functions that human-operated businesses assign to management. These are AI-level judgment tasks that require a different kind of agent, not a different kind of person.

The Agent Council performs four specific functions. Quality review: each execution agent output is evaluated against established Quality Threshold standards before delivery, checked against the Operational Ledger and known failure modes for the task class — at compute cost, within the same execution cycle, without a queue. Error detection: outputs are cross-referenced against the Proof of Action trail using a contradiction protocol that does not share the execution agents' failure modes, because an Agent Council that shares the same model and context provides corroborated confirmation, not independent review. Escalation triage: the council resolves what it can and surfaces only conditions exceeding its resolution authority to the Audit Surface. Pattern learning: the council identifies recurring failure patterns across executions and surfaces improvement signals before they compound into Logic Decay.

The independence requirement is the most critical architectural specification. Effective council-layer review requires a different evaluation angle from the execution layer: a structured contradiction protocol, a different model, or a methodology that actively looks for inconsistency rather than confirming correctness. Methodological independence is the source of the council's value as a review layer.

The Agent Council is also the architectural resolution of the Authorization Trap at the quality review checkpoint: once the council demonstrates through a supervised operation window that its error detection meets the promotion criteria, the human approval step above it becomes an undepurated Human Premium component rather than a safety measure. The council must be specified at Full-System Design time — its pattern-learning function compounds from the first execution cycle, and retrofitting it into an operating function means reconstructing baselines from a history not designed to produce them.

First used: July 2026

Edition 1 · updated July 2026