ACGP Concepts Overview

ACGP adds runtime governance to an agent system without requiring governance logic to live inside each agent. It standardizes how deployments describe policy, evaluate proposed actions, and preserve audit-visible outcomes.

What ACGP Adds To An Agent System

ACGP adds a governance layer between an operating agent and its real-world effects.

The canonical runtime sequence is tripwires -> deterministic checks -> CTQ -> thresholds -> flag -> trust debt -> governance-tier review -> audit. Full operational ordering and edge-case semantics are defined in ACGP-3, with tripwire fail-closed rules in ACGP-4.

At a high level, that means:

  • agents still propose actions and produce reasoning or evidence
  • policy is defined outside the agent in portable artifacts
  • runtime governance evaluates the proposed action before execution
  • interventions and audit artifacts remain visible after the decision

Main Actors And Artifacts

The core moving parts are:

  • Operating Agent: the system proposing actions, tool calls, or workflow steps.
  • Governance Steward: the runtime authority that evaluates governed actions and emits outcomes.
  • Reflection Blueprint: the policy artifact that defines checks, thresholds, evidence requirements, and trust-policy behavior.
  • Cognitive Trace: the structured reasoning and action evidence presented for evaluation.
  • Intervention: the runtime outcome that decides whether the action proceeds, is modified, is escalated, or is denied.
  • Governance Store / audit trail: the durable record of evidence, intervention, and review-visible outcomes.
  • Governance Tier: the ARS-derived default strictness posture applied by the deployment.

Governance Flow

flowchart TD
    A["Action proposed"] --> B["Cognitive Trace observed"]
    B --> C["Tripwires and deterministic checks"]
    C --> D["CTQ and threshold evaluation"]
    D --> E["Intervention decided"]
    E --> F["Trust debt and audit effects recorded"]
    F --> G["Human review or stronger posture where required"]

Profile-failure fallback handles Steward/session-path unavailability. Evaluation-timeout policy handles an exceeded negotiated latency budget while the Steward/session path remains available.

Step-by-step

  1. An action is proposed.
  2. A Cognitive Trace is observed.
  3. Tripwires, deterministic checks, and evaluation are applied.
  4. An intervention is decided.
  5. Trust and audit effects are recorded.
  6. Human review or stronger posture may apply where required.

For the full operational ordering, see ACGP-3 and ACGP-4.

What Is Core In Alpha Vs Preview Vs Deployment-Defined

  • Core in alpha: Reflection Blueprints, Cognitive Traces, tripwires, CTQ, threshold mapping, interventions, Trust Debt observables, Governance Tier posture, and durable audit semantics.
  • Preview in alpha: Runtime Governance Contracts and other extension-track material remain non-core unless a future extension suite activates them.
  • Deployment-defined: identity proof, key lifecycle, trust bootstrap, federation, topology, and remote assurance controls remain deployment responsibilities.

See Protocol Boundaries for the conceptual boundary map.

Intervention Types

ACGP uses five primary intervention levels plus orthogonal flag:

  • ok: action proceeds.
  • nudge: action proceeds with guidance.
  • escalate: human review is required before execution continues.
  • block: action is denied while the session remains active.
  • halt: emergency stop; the active session is terminated.
  • flag (orthogonal): adds audit and review tagging plus trust-debt impact to any primary intervention.

Tripwires run before CTQ threshold mapping and can force block or halt regardless of score.

How Teams Adopt ACGP Incrementally

  • Start with one steward, one active baseline blueprint, and one durable audit path.
  • Add Governance Tier assessment so runtime posture is explicit instead of ad hoc.
  • Introduce stronger evidence requirements, tripwires, and review posture where the operating envelope demands it.
  • Add preview extensions only when the deployment actually needs them.

This overview page owns the component map and adoption model; the walkthrough and specs own the deeper runtime sequence details.

Next Steps

  • Read How ACGP Works for a guided walkthrough of one governed action.
  • Review Protocol Boundaries for deployment responsibility and trust-root boundaries.
  • Continue into ACGP-3 for the normative operational model.