Pilot
AI Agent Safety Pilot
Start with one real action your team cares about. Sanctum verifies it at runtime, shows the operator what is at stake, and records proof before the side effect runs.
What the pilot proves
Sanctum is not another prompt rule. It is a runtime decision point between agent intent and execution. In a pilot, your team sees an agent propose an action, Sanctum evaluates policy and risk, and execution only continues when the decision allows it.
The outcome is simple: your AI agent can act, but not without runtime permission.
Best-fit teams
- AI SaaS teams whose agents send messages, update CRMs, deploy code, or change customer data
- Platform teams adding MCP, LangChain, CrewAI, Vercel AI SDK, or OpenAI tool calling to production
- Security and compliance teams that need approval evidence before expanding agent permissions
- Robotics, smart-home, and industrial teams moving from demos to physical-world autonomy
The five-minute path
- Create or select a Sanctum agent in the console.
- Choose Connect Agent for the fastest proxy path, or the SDK for code-level control.
- Route one realistic tool call through the runtime.
- Review the hold, block, or approval in Live Feed with source trust, blast radius, and policy reason.
- Approve, deny, or convert the observed tool into a standing policy.
Actions to test first
Start where a bad action would be expensive, visible, irreversible, or hard to explain.
send_customer_email
transfer_funds
delete_database_record
deploy_production
read_secret
unlock_door
change_robot_mode
export_phi
What good looks like
- Control: risky actions are held or blocked before execution.
- Context: operators see actor, tool, source trust, blast radius, and policy reason.
- Proof: audit logs show who approved, denied, or changed policy.
- Enforcement: executors can require a signed action token before running.
- Expansion: the same pattern scales from one tool to a fleet.
Full reference: documentation · llms.txt · architecture.md