Agent Intent Protocol Documentation¶

Welcome to the AIP Kubernetes Control Plane documentation. These guides explain how to govern autonomous AI agents operating on critical infrastructure.

What is AIP?¶

The Agent Intent Protocol is an open standard that requires autonomous agents to declare their intentions before acting. AIP decouples agent decision-making from system safety, giving operators:

Policy enforcement — CEL-based rules that evaluate every request
Human-in-the-loop gates — approval workflows for high-risk actions
Audit trails — immutable records of every state transition
Earned autonomy — agents graduate from Observer → Autonomous based on measured accuracy

Getting started¶

Install the control plane in under a minute:

helm install aip-k8s \
  oci://ghcr.io/agent-control-plane/aip-k8s/charts/aip-k8s \
  --version 0.1.0 \
  --namespace aip-k8s-system \
  --create-namespace

Then follow the Quick Start to submit your first governed request and see AIP block an agent operating on stale data.

⚠️ Default install is dev mode — no authentication. See Dev Mode and Production Hardening before deploying to a real cluster.

Core concepts¶

Guide	What you'll learn
Agent Graduation Ladder	How agents progress from Observer to Autonomous through measured accuracy
Trust Gate	How the gateway enforces trust levels on every request
Governed Resources	How to register infrastructure resources that agents may target
Garbage Collection	How to manage retention and export of diagnostic records

User guide¶

Guide	What you'll learn
Quick Start	Install, verify, and submit your first request
Dashboard Walkthrough	Visual inspection of requests, policy decisions, and audit trails
Dev Mode	What the default install enables and disables
Production Hardening	OIDC, roles, JWT, and security checklist
Scaledown Demo	End-to-end: agent on stale metrics gets blocked

Authentication¶

Guide	What you'll learn
OIDC with Keycloak	Configure JWT authentication with Keycloak (recommended for local dev)

Quick reference¶

The five trust levels¶

Level	Execution	Human approval
Observer	Graded only, no action	N/A (grading)
Advisor	Allowed	Required
Supervised	Allowed	Required
Trusted	Allowed	Auto-approved by policy
Autonomous	Allowed	Auto-approved by policy

Key CRDs¶

AgentRequest — an agent declares intent to act on a target resource
GovernedResource — platform engineering registers permitted resource types
SafetyPolicy — CEL rules that further restrict what agents may do
AgentGraduationPolicy — cluster admin defines accuracy thresholds per level
AgentTrustProfile — controller-managed record of an agent's earned trust
DiagnosticAccuracySummary — rolling accuracy computed from graded verdicts
AuditRecord — immutable event log for every state transition

Resource access patterns: GovernedResource and SafetyPolicy are managed through the gateway REST API as the primary path (kubectl is break-glass only). AgentGraduationPolicy, AgentTrustProfile, DiagnosticAccuracySummary, and AuditRecord have no gateway endpoints yet — inspect them with kubectl. See the Trust Gate guide for details.