Stop Your AI Agent Before It Makes a $50,000 Mistake

Your AI agent is about to issue a refund. Should it?

Request Amount ACGP Decision Time
Defective product return $50 OK - Approved 12ms
Suspicious bulk refund $50,000 ESCALATE - Human review 45ms

ACGP checks every agent decision in real-time. Risky actions get caught. Safe ones flow through.

from acgp import GovernanceSteward, CognitiveTrace

steward = GovernanceSteward(blueprint_file="blueprint.yaml")
trace = CognitiveTrace(
    reasoning="User requested refund for defective item",
    action="issue_refund",
    parameters={"amount": 50}
)
result = steward.evaluate(trace)
print(result.intervention)  # "OK" - approved in 12ms

Try It Now How It Works


What is ACGP?

The Agentic Cognitive Governance Protocol is an open specification for governing autonomous AI agents at runtime. It provides:

  • Governance Contracts: Performance SLAs, risk levels (low/elevated/critical), evaluation tiers (Eval-0 to Eval-3), and fallback behaviors (Standard+ conformance)
  • Cognitive Traces: Transparent reasoning trails with versioning and attribution
  • Interventions: Six intervention types (five primary levels: OK, Nudge, Escalate, Block, Halt, plus orthogonal Flag)
  • Tripwires: Proactive safeguards against resource abuse and drift
  • Trust System: Gradual autonomy through proven behavior
  • Reflective Learning: Continuous improvement from outcomes

ACGP is designed to work with any agent framework (LangChain, AutoGPT, custom) and provides three conformance tiers:

  • Minimal: Development and batch operations (2-4 hours implementation)
  • Standard: Production-ready governance with SLAs (1-2 days)
  • Complete: Mission-critical systems with HSM and consensus (1-2 weeks)

Core Concepts

ARI Framework

The Agent-Reflection-Intervention framework is the heart of ACGP:

graph LR
    A[Agent Decision] --> R[Governance Steward Reflection]
    R --> CTQ{Cognitive Trace Quality OK?}
    CTQ -->|Yes| OK[ Proceed]
    CTQ -->|Concerns| I[Intervention]
    I --> N[Nudge/Flag/Escalate/Block]
  • Agent: Makes decisions and takes actions
  • Reflection: Governance Steward evaluates decision quality
  • Intervention: Proportionate response based on concerns

Learn about ARI

ACL Tiers

Access Control Level (ACL) tiers determine how strictly agents are governed:

ACL Tier ARS (Agent Risk Score) Use Case Latency
ACL-0 0-2 Read-only bots ~10ms typical, <50ms max
ACL-1 3-4 Scripted agents ~20ms typical, <100ms max
ACL-2 5-7 Moderate autonomy ~50ms typical, <150ms max
ACL-3 8-10 High autonomy ~100ms typical, <200ms max
ACL-4 11-13 Critical systems ~200ms typical, <350ms max
ACL-5 14-15 Maximum oversight ~500ms typical, <1000ms max

Understand ACL tiers

Interventions

Six intervention types (five primary levels: OK, Nudge, Escalate, Block, Halt, plus orthogonal Flag):

  • OK - Continue without interruption
  • Nudge - Gentle suggestion to reconsider
  • Escalate - Request human review
  • Block - Prevent this specific action
  • Halt - Stop agent completely
  • Flag - Log for review (can combine with any primary level)

Note: Flag is orthogonal, meaning it can be combined with any of the five primary intervention levels for audit and trust debt tracking.

Intervention types


Example: Customer Service Agent

from acgp import GovernanceSteward, CognitiveTrace

# Initialize steward for customer service agent
steward = GovernanceSteward(
    blueprint_file="customer-service-v1.yaml",  # (1)!
    conformance_level="standard",  # (2)!
)

# Agent makes a decision
trace = CognitiveTrace(
    reasoning="Customer has valid complaint, offering 15% discount",
    action="apply_discount",
    parameters={"amount": 15, "reason": "service_recovery"}
)

# Evaluate with ACGP
result = steward.evaluate(trace)

if result.intervention == "OK":
    # Apply the discount
    apply_discount(15)
elif result.intervention == "NUDGE":
    # Log suggestion but proceed
    log_suggestion(result.message)
    apply_discount(15)
elif result.intervention == "BLOCK":
    # Don't apply, inform user
    notify_agent(result.message)
  1. Blueprints define governance policies specific to your use case. Create custom blueprints with the Blueprint Schema.
  2. Standard conformance provides production-ready governance with governance contracts and SLAs.

See more examples


ACGP vs. Alternatives

Feature Traditional IAM Rate Limiting Monitoring ACGP
Real-time prevention [NO] [YES] [NO] [YES]
Behavioral analysis [NO] [NO] [WARNING] [YES]
Graduated response [NO] [NO] [NO] [YES]
Quality evaluation [NO] [NO] [WARNING] [YES]
Learning/adaptation [NO] [NO] [NO] [YES]
Compliance auditing [WARNING] [NO] [YES] [YES]
Framework agnostic [YES] [YES] [YES] [YES]

Tip

Use ACGP in addition to IAM, rate limiting, and monitoring—not as a replacement. ACGP provides a behavioral governance layer that complements traditional security controls.


When to Use ACGP

Perfect for:

  • Autonomous agents that make decisions without constant human oversight
  • High-risk actions like financial transactions, data modifications, or external communications
  • Adaptive AI that learns and changes behavior over time
  • Multi-agent systems that need coordinated governance
  • Compliance requirements that mandate decision audit trails

Not needed for:

  • Simple scripted bots with hard-coded if/else logic
  • Read-only agents that never take actions
  • Sub-10ms latency requirements (ACGP adds 10-500ms depending on ACL tier)
  • Non-cognitive systems like traditional automation

Based on community usage:

  1. Getting Started - Get started in 5 minutes
  2. ARS Calculator - Assess your agent
  3. Intervention Types - Understand responses
  4. Troubleshooting - Fix common issues
  5. MCP Integration - Connect to tools

Implementation Paths

Choose your implementation level based on your needs:

Compare conformance levels


Specifications

ACGP is defined through a series of RFC-style specifications:

Browse all specifications


Community


Next Steps