What is ACGP?¶

The Agentic Cognitive Governance Protocol (ACGP) provides runtime governance for AI agents through continuous behavioral evaluation.

The Problem¶

AI agents are becoming increasingly autonomous, but traditional access control systems aren't designed for cognitive systems that:

Make complex decisions based on reasoning
Adapt their behavior over time
Operate with varying degrees of autonomy
Take actions with real-world consequences

Traditional solutions fall short:

IAM - Controls what agents can access, not how they reason
Rate limiting - Prevents abuse, not poor decision-making
Monitoring - Detects problems after they happen, not before

The ACGP Solution¶

ACGP evaluates the quality of agent reasoning in real-time, enabling proportionate interventions before actions are executed.

graph TD
    A[Agent Decision] --> C[Cognitive Trace]
    C --> S[Governance Steward]
    S --> E{Evaluation}
    E -->|Good Quality| OK[ OK - Proceed]
    E -->|Minor Concerns| N[ Nudge - Suggest Improvement]
    E -->|High Risk| ES[ Escalate - Request Human]
    E -->|Unacceptable| B[ Block - Prevent Action]
    E -->|Critical Violation| H[ Halt - Stop Agent]

    F[ Flag - Log for Review] -.->|Orthogonal: can combine with any| OK
    F -.-> N
    F -.-> ES

Key Features¶

Continuous Evaluation¶

Every agent decision is evaluated before execution, providing real-time governance.

Graduated Responses¶

Six intervention types: five primary levels (OK, Nudge, Escalate, Block, Halt) plus orthogonal Flag that can combine with any primary level.

Dynamic Trust¶

Agents earn autonomy through consistent good behavior, lose it through poor decisions.

Framework Agnostic¶

Works with any agent architecture - wrap existing agents without modification.

Three Conformance Levels¶

Choose Minimal (dev), Standard (prod), or Complete (mission-critical) based on your needs.

How It Works¶

1. Cognitive Traces¶

Agents create traces of their reasoning and planned actions:

{
  "reasoning": "User requested password reset. Email verified, account locked due to suspicious activity.",
  "action": "send_password_reset_email",
  "parameters": {
    "email": "user@example.com",
    "include_unlock": true
  }
}

2. Governance Evaluation¶

The Governance Steward evaluates trace quality based on:

Reasoning quality - Is the logic sound?
Action appropriateness - Is this the right action?
Policy compliance - Does it follow rules?
Risk assessment - What could go wrong?

3. Interventions¶

Based on evaluation, the steward issues an intervention type:

Five Primary Levels:

OK - Good to go
Nudge - Consider alternative
Escalate - Get human approval
Block - Don't do this
Halt - Stop everything

Plus Orthogonal Flag:

Flag - Log for review (can combine with any primary level)

When to Use ACGP¶

Perfect For¶

Autonomous agents making unsupervised decisions
High-risk actions (financial, data, external communications)
Adaptive systems that learn and change
Multi-agent systems needing coordinated governance
Compliance requirements demanding audit trails

Not Needed For¶

Simple scripted bots (if/else logic)
Read-only agents (no actions)
Sub-10ms latency requirements
Non-cognitive automation

Core Concepts¶

ARI Framework

Agent-Reflection-Intervention: The core governance model
ACL Tiers

Six levels of oversight from minimal to maximum
Cognitive Traces

Capturing agent reasoning for evaluation
Interventions

Six graduated response levels
Tripwires

Hard limits that trigger immediate blocks
Trust System

Dynamic trust scoring and adjustment