Why ACGP Exists

The Problem: Autonomous Agents Can Go Wrong

AI agents are increasingly making autonomous decisions—from customer service to financial transactions to infrastructure management. But what happens when they make mistakes?


Real Incidents (Anonymized)

Incident 1: The $10K AWS Bill

What happened: - Autonomous agent had AWS EC2 access for scaling infrastructure - Bug in cost estimation logic triggered runaway scaling - Spun up 200 GPU instances in 2 hours - Developer noticed after getting AWS billing alert - Cost: $10,000 in 2 hours

Root cause: No budget tripwire, no escalation for unusual spending patterns

How ACGP prevents this:

tripwires:
  - id: "daily_aws_spending"
    condition: "daily_total > 1000"
    action: "HALT"
    message: "Daily AWS spending limit exceeded"


Incident 2: The Refund Bot

What happened: - Customer service AI had authority to issue refunds up to $500 - Missing upper limit validation in production deployment - Approved 90% of refund requests over 2 weeks - Average refund: \(250 (baseline was ~\)50) - Cost: $150,000 before detection

Root cause: No anomaly detection, overly permissive policy, no escalation

How ACGP prevents this:

policies:
  - action: issue_refund
    if: "amount > 100"
    then: "ESCALATE"
    message: "Refunds over $100 require manager approval"

tripwires:
  - id: "daily_refund_total"
    condition: "daily_refunds > 2000"
    action: "HALT"


Incident 3: The Trading Algorithm

What happened: - Algorithmic trading bot with autonomous execution - Flash crash triggered failsafe bug - Executed 10,000 panic sells in 45 seconds - Market order slippage compounded losses - Loss: $2.3M in 45 seconds

Root cause: No circuit breaker, no human-in-the-loop for massive positions

How ACGP prevents this:

governance_contract:
  risk_level: "critical_risk"
  performance_budget:
    latency_budget_ms: 1000
    fallback_behavior: "deny"

rules:
  - action: "execute_trade"
    if: "position_size > 100000"
    then: "ESCALATE"
    requires: "human_approval"


The Solution: Cognitive Governance

ACGP provides runtime governance that evaluates agent decisions before they execute, using:

1. Automatic Budget Tripwires

Problem: Agents can rack up costs before anyone notices
Solution: Real-time budget tracking with automatic halt

tripwires:
  - when: daily_spend > 1000
    then: HALT
  - when: hourly_api_calls > 10000
    then: ESCALATE

Result: Catch runaway costs in seconds, not hours


2. Action Limits by Risk

Problem: One-size-fits-all permissions are too coarse
Solution: Graduated interventions based on action risk

policies:
  - action: issue_refund
    if: amount <= 50
    then: OK
  - action: issue_refund
    if: 50 < amount <= 500
    then: ESCALATE
  - action: issue_refund
    if: amount > 500
    then: BLOCK

Result: Low-risk actions proceed instantly, high-risk actions get review


3. Complete Audit Trail

Problem: Can't reconstruct what happened or why
Solution: Every decision logged with reasoning

trace = CognitiveTrace(
    reasoning="Customer reported defective product...",
    action="issue_refund",
    parameters={"amount": 250, "order_id": "12345"}
)

result = steward.evaluate(trace)

# Logged to ReflectionDB:
# - Full reasoning chain
# - Decision and intervention
# - CTQ score and risk assessment
# - Human override (if escalated)
# - Timestamp and session context

Result: Full forensic capability for compliance and debugging


4. Gradual Autonomy

Problem: New agents are untested but need autonomy
Solution: Earn trust through good behavior

# New agent starts with low trust (ACL-2)
initial_acl = "ACL-2"  # More oversight, higher latency

# After 30 days of good behavior:
# - 0 interventions
# - CTQ score consistently >0.85
# - No tripwire violations

# Auto-promoted to ACL-3
current_acl = "ACL-3"  # More autonomy, lower latency

Result: Risk decreases as agents prove themselves


What ACGP Prevents

Risk Without ACGP With ACGP
Runaway costs Detected in billing cycle (days/weeks) Halted in seconds via tripwires
Policy violations Discovered in audit (after damage) Prevented before execution
Anomalous behavior Manual review of logs Automatic flagging and escalation
Catastrophic errors Hope and pray Circuit breakers and HALT
Compliance gaps Reconstruct from logs (maybe) Complete audit trail always

When You Need ACGP

Perfect for:

  • Autonomous agents making decisions without constant supervision
  • High-risk actions like financial transactions, data deletion, privilege grants
  • Adaptive AI that learns and changes behavior
  • Compliance requirements needing decision audit trails
  • Multi-agent systems requiring coordinated governance

Consider carefully:

  • Latency-sensitive operations (<10ms requirements)
  • Read-only agents with no side effects
  • Simple scripted bots with hard-coded logic

Not needed:

  • Static automation with no decision-making
  • Systems with existing robust human oversight
  • Prototypes with no real-world consequences

Real-World Use Cases

Customer Service Automation

  • Risk: Agents approve refunds, discounts, policy exceptions
  • ACGP Solution: Escalate refunds >\(100, halt if daily total >\)5000
  • Impact: Reduced fraud by 87%, caught 12 policy violations in first month

Infrastructure Management

  • Risk: Agents scale resources, modify configurations, deploy code
  • ACGP Solution: Block production changes without approval, halt on unusual spending
  • Impact: Prevented 3 potential outages, saved $50K in cloud costs

Financial Trading

  • Risk: Agents execute trades, adjust positions, manage risk
  • ACGP Solution: Escalate large positions, deny on margin calls, circuit breakers
  • Impact: Zero incidents in 18 months vs 4 incidents in prior year

Data Operations

  • Risk: Agents query databases, export data, modify schemas
  • ACGP Solution: Flag large exports, block production modifications, escalate privilege changes
  • Impact: 100% compliance with data privacy regulations

Getting Started

Ready to add governance to your agents?

  • 5-Minute Start


    Add basic governance in 5 minutes with Minimal conformance

    Quick Start

  • Production Deploy


    Full governance with contracts, tripwires, and audit trails

    Standard Conformance

  • Calculate Latency


    Plan your governance contracts and latency budgets

    Latency Calculator

  • Learn Concepts


    Understand ARI framework, ACL tiers, and interventions

    Core Concepts