Agents
Registering and governing AI Agents
In ChooChoo, an Agent is a first-class passenger. It is an actor that can produce artifacts, execute workflows, and make decisions. Every agent action is recorded in the Audit Trail, and agents are governed by boundaries that the validation engine enforces.
Agent Registration
All agents operating within a governed repository MUST be registered in agents/AGENTS.md. This file serves as a human-readable registry and a machine-readable configuration source. See the Project Structure for where this fits in the directory layout.
Format
# Registered Agents
## Cursor Agent 001
**ID:** `cursor-agent-001`
**System Card:** `system-cards/cursor.yaml`
**Status:** `active`
### Capabilities
- Read codebase
- Propose code changes
- Run tests
### Boundaries
- Cannot modify `contracts/` without approval
- No access to `production` environment
- Cannot read fields tagged `pii`Each agent should have a corresponding System Card that documents its capabilities, limitations, and compliance frameworks. System Cards are referenced during risk scoring to determine the level of trust assigned to an agent.
Boundaries (Rails)
Boundaries defined in the Agent Registry are enforced by the ChooChoo Engine. If an agent attempts an action that violates a boundary, the CLI will block the operation and return error E007 (Agent boundary violation).
| Boundary Type | Description |
|---|---|
read-only | Agent cannot modify artifacts of specific types. |
no-pii | Agent receives redacted values for PII fields. See Security Considerations. |
requires-approval | Agent actions trigger a human approval workflow. |
env-restricted | Agent cannot operate in specific environments (e.g., Prod). Affects risk scoring. |
scope-limited | Agent is restricted to specific file paths. |
Boundary enforcement is part of the validation pipeline. In CI/CD, boundary violations produce exit code 1 (VALIDATION_ERROR).
Agent Activity Log
Every action an agent takes is recorded in the Audit Trail. This log links the agent to its System Card, the artifact modified, and the full decision context captured via Agent Trace.
{
"timestamp": "2026-01-31T14:23:45Z",
"agentId": "cursor-agent-001",
"action": "modify",
"artifact": "contracts/customer-360.contract.yaml",
"context": {
"prompt": "Add phone number field",
"confidence": 0.85
}
}The confidence field feeds into the risk scoring algorithm as the Agent Confidence (A) factor — lower confidence increases the risk score.
Managing Agents via CLI
Use the choochoo agent commands to manage your agent registry:
choochoo agent list # List all registered agents
choochoo agent register # Interactively register a new agent
choochoo agent audit <id> # View activity history for an agentSee the CLI Reference for full documentation on all agent management commands.
Related
Agent Trace
Learn how AI context is persisted and linked to code changes for auditability.
System Cards
Create identity documents that define an agent's capabilities and compliance posture.
Audit Trail
Explore the immutable log of all agent and human decisions.
Risk Scoring
Understand how agent confidence and boundaries affect the risk calculation.
Approval Workflows
Define policy gates that require human sign-off for agent actions.
Security Considerations
PII handling, encryption requirements, and audit log integrity for agent access.