ACGP-0000: Protocol Overview (START HERE)

Status: Draft
Last Updated: 2026-01-08
Spec ID: ACGP-0000
Normative Keywords: MUST, SHOULD, MAY (per RFC 2119)


READ THIS FIRST

This document is your entry point to the Agentic Cognitive Governance Protocol (ACGP). It provides a high-level overview and guides you to the right detailed specifications based on your role and needs.


TL;DR (30 Seconds)

ACGP provides runtime governance for AI agents, ensuring they operate safely and align with organizational policies through continuous cognitive quality monitoring.

How it works:

  1. Assess agent risk - Assign control level (ACL tier)
  2. Configure governance policies via Reflection Blueprints
  3. Monitor every agent action via Cognitive Traces
  4. Evaluate action quality (CTQ score)
  5. Intervene based on risk (OK, Nudge, Escalate, Block, Halt, Flag)

5-Minute Overview

What Problem Does ACGP Solve?

AI agents are becoming more autonomous and capable, but without governance they can:

  • Make unsafe decisions
  • Access unauthorized resources
  • Generate harmful content
  • Operate outside policy boundaries
  • Lack accountability and auditability

ACGP solves this by providing real-time oversight without requiring changes to your agent's core logic.

Core Architecture

┌─────────────────────────────────────────────────────────────┐
│  Layer 1: RISK ASSESSMENT (Static/Periodic)                 │
│  ┌──────────────┐         ┌─────────────────┐               │
│  │ ARS          │────────>│  ACL Tier       │               │
│  │ (0-15)       │         │  (ACL-0 to 5)   │               │
│  └──────────────┘         └─────────────────┘               │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│  Layer 2: POLICY CONFIGURATION                              │
│  ┌──────────────────────┐  ┌─────────────────────────┐      │
│  │ Reflection           │  │ Certified Source        │      │
│  │ Blueprints           │  │ Registry                │      │
│  └──────────────────────┘  └─────────────────────────┘      │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│  Layer 3: RUNTIME EVALUATION (Continuous)                   │
│                                                             │
│  Agent Action → Cognitive Trace → CTQ Evaluation            │
│                                           ↓                 │
│                                    Intervention Decision    │
│                                    (OK/Nudge/Flag/etc)      │
└─────────────────────────────────────────────────────────────┘

Key Concepts

ARS (Agent Risk Score): Numerical score (0-15) assessing agent risk across three dimensions

  • Evaluates: Autonomy, Adaptability, Continuity (each 0-5)
  • Total Score: Sum of all three dimensions (0-15)
  • Determines: Which ACL tier the agent operates under

ACL (Access Control Level): Risk-based control tiers (ACL-0 to ACL-5)

  • Higher risk = Stricter governance
  • Defines: Decision thresholds and required safeguards

Cognitive Trace: Structured log of agent's reasoning and actions

  • Contains: Inputs, reasoning, tool calls, outputs
  • Purpose: Enable quality evaluation

CTQ (Cognitive Trace Quality): Runtime quality score (0-1)

  • Five metrics: Reasoning Quality, Knowledge Grounding, Ethical Alignment, Tool Safety, Context Awareness
  • Higher score = Better quality, lower risk

Intervention: Governance decision based on CTQ score

  • Six types (five primary levels: OK, Nudge, Escalate, Block, Halt, plus orthogonal Flag)
  • Proportionate response to detected risk

Who Should Read What?

I'm New to Agent Governance

Start here:

  1. Read this document (ACGP-0000) - You're here!
  2. Read ACGP-1001: Terminology - Understand key terms
  3. Read ACGP-1000: Core Protocol - Understand the basics
  4. Try: Quick Start Tutorial (coming soon)

I'm Implementing ACGP

Read in this order:

  1. ACGP-1000: Core Protocol - Foundation
  2. ACGP-1002: Architecture - System design
  3. ACGP-1003: Message Formats - Wire protocol
  4. ACGP-1005: ARS-CTQ-ACL Integration Framework - Evaluation logic
  5. ACGP-1009: Conformance - Testing requirements

I'm Integrating with Existing Agents

Focus on:

  1. ACGP-1008: Interoperability - MCP/A2A integration
  2. ACGP-1004: Blueprints - Policy configuration
  3. SDK Documentation (coming soon)

I'm Focused on Security

Start with:

  1. ACGP-1007: Security Considerations - Threat model
  2. ACGP-1006: Certified Source Registry - Trust anchors
  3. ACGP-1005: Section 6 - Tripwires

I'm Evaluating Compliance

Read:

  1. ACGP-1009: Conformance Requirements
  2. ACGP-1005: ARS-CTQ-ACL Framework
  3. Conformance Checklist (Appendix A of ACGP-1009)

Document Map

ACGP-0000 (Overview) <-- YOU ARE HERE
    |
    +-- ACGP-1000 (Core Protocol) <-- Start here for basics
    |       |
    |       +-- ACGP-1001 (Terminology)
    |       +-- ACGP-1002 (Architecture)
    |       +-- ACGP-1003 (Message Formats)
    |       +-- ACGP-1004 (Blueprints)
    |       +-- ACGP-1005 (ARS-CTQ-ACL Framework)
    |
    +-- ACGP-1006 (Certified Source Registry)
    +-- ACGP-1007 (Security Considerations)
    +-- ACGP-1008 (Interoperability)
    +-- ACGP-1009 (Conformance Requirements)

Quick Reference: Key Numbers

Concept Range Meaning
ARS (Agent Risk Score) 0-15 Agent risk level (higher = more risk)
ACL Tier 0-5 Control strictness (higher = stricter)
CTQ Score 0-1 Quality score (higher = better)
Risk Score 0-1 Inverse of CTQ (higher = worse)

ARS to ACL Mapping

ARS (Agent Risk Score) ACL Tier Risk Level
0-2 ACL-0 Minimal
3-4 ACL-1 Low
5-7 ACL-2 Medium
8-10 ACL-3 High
11-13 ACL-4 Very High
14-15 ACL-5 Critical

Intervention Types

Intervention Action When Used
OK Allow and log CTQ is good, no issues detected
Nudge Modify and proceed Minor issues, auto-correctable
Flag Allow but audit Concerning but not blocking
Escalate Pause for human review Uncertain, needs judgment
Block Prevent this action Clear policy violation
Halt Kill switch Critical safety issue

Implementation Levels

ACGP defines three conformance levels:

Minimal Conformance

Purpose: Development and testing
Requirements: Core protocol, basic governance, local storage
Use Cases: Development environments, proof of concept

Standard Conformance

Purpose: Production deployment
Requirements: All Minimal + interoperability, distributed storage, security
Use Cases: Enterprise production systems

Complete Conformance

Purpose: Mission-critical systems
Requirements: All Standard + advanced security, high availability
Use Cases: High-stakes, regulated environments

See ACGP-1009 for detailed requirements.


Common Questions

Q: How does ACGP differ from traditional access control?

A: Traditional access control is binary (allow/deny) and static. ACGP provides continuous evaluation with graduated interventions and adaptive policies based on runtime behavior.

Q: Does ACGP require changes to my agent code?

A: Minimal. You add a wrapper that emits Cognitive Traces. The governance logic runs separately in the Governance Steward.

Q: Can different vendors implement ACGP?

A: Yes! The protocol is open. Vendors can compete on implementation quality while maintaining interoperability through standardized message formats.

Q: What if my agent doesn't fit the ARS rubric?

A: The ARS assessment is extensible. You can add custom dimensions or use the manual override process.

Q: How much latency does ACGP add?

A: Target: <100ms for ACL-0/1, <200ms for ACL-⅔, <500ms for ACL-⅘. See ACGP-1002: Section 9 for details.


Getting Started Checklist

  • Read this overview document (you're almost done!)
  • Review the Terminology
  • Understand the Core Protocol
  • Choose your implementation level
  • Read relevant specifications for your role
  • Review conformance requirements
  • Check out reference implementation (coming soon)
  • Join the community (links coming soon)

Contributing

ACGP is an open protocol. We welcome:

  • Implementation feedback
  • Specification improvements
  • Use case documentation
  • Reference implementations
  • Integration adapters

Repository: [Coming soon]
Discussion: [Coming soon]
Issues: [Coming soon]


License

[To be determined - recommend permissive open source license]


Next Steps

Ready to dive deeper?


For questions, feedback, or contributions, please [contact us / open an issue / join our community]