From Agent Prototype to Safe Operations
How do you move a promising AI agent into a safely operated system?
About 25 minutes · 6 stops · progress stays in this browser.
Field journal
Define what the agent can actually do
Classify tools, autonomy, state, delegation, and environmental risk before selecting infrastructure.
Map the managed-service boundaries
Separate runtime, identity, gateway, memory, observability, and the controls the application still owns.
Make deployment reviewable
Turn agent delivery into a deterministic platform workflow with policy, evaluation, approval, and rollback.
Enforce trust at every hop
Design workload identity, user delegation, narrow authority, audit, and recovery around consequential actions.
Prove state migration and rollback
Use a real Hermes migration to see why writer shutdown, SQLite validation, single-consumer cutover, and rollback matter.
Measure the swarm before you trust its policy
Use a 33-run surf-school swarm to see the completion rate, tail latency, cost, and failure modes a clean architecture diagram hides, and why hard constraints belong in code.
Design the operating boundaries, not just the prototype
Principles collected
- Describe authority and risk before calling a system an agent.
- Treat managed services as boundaries, not as automatic production readiness.
- Make infrastructure changes deterministic, reviewable, and reversible.
- Authenticate and authorize every workload and delegated user at each hop.
- Test state recovery and cutover behavior with the actual runtime.
- Enforce hard constraints in deterministic code, not in prompt sentences.
Practical checklist
- Record the agent's tools, autonomy, state, and maximum consequence.
- Separate inbound authentication, workload identity, and tool authorization.
- Pin artifacts and require evaluation, approval, observability, and rollback evidence.
- Minimize network reach and permissions for every component.
- Rehearse backup, restore, writer shutdown, and rollback before migration.
- Log every hop (node path, tokens, tool calls, failures) and budget the tail, not the average.