Skill · production · durable-execution

Durable execution

Run the agent as a reducer over one serializable blob with launch/pause/resume, idempotent tools, and crash replay. Use when a customer will wait, HITL must survive, Inngest/LangGraph checkpoints, or the demo works on a laptop.

GET /api/canon/skills/durable-execution?format=md

Factors 5, 6, 12. The ticket is the agent state. Crashes replay. Humans resume. Double side effects are a product incident.

When

A demo works on your laptop and a customer will wait on it.

Do

01
One blob
Business fields and loop fields together. Reconstruct the next LLM call from the blob plus pinned prompts.
02
Boring APIs
Launch, pause, resume. Interrupts are first-class. Gateway from mail/chat/cron/webhook into the same reducer.
03
Idempotent tools
Keys, journals. 'Already refunded' is visible. Retries do not charge twice.

Don't

A second agent_memory table that drifts from the ticket.
Restart the agent after the human answers.

Hard rules

(state, event) → (newState, effects).
Pause serializes; resume does not re-execute.
One orchestrator.

Refuse

HITL that cannot resume — The human answers and the run is gone. Approval as a side channel, not a tool.
Double side effects — Non-idempotent tools, no journal. A retry refunds twice.
Unbounded loop — No max-turns, token, wall-clock, or money budget. The agent rambles until the bill or the context dies.

Load with this

Job · Production agent product Job · Support / ops agent Playbook · Ship durable Playbook · Add a human gate Durable Execution The Agent as a Stateless Reducer Unify Execution and Business State Observability Trigger from Anywhere

Load next

Trigger tests

Should fire

“The agent dies if the server restarts”
“Durable graph with HITL”
“Inngest vs LangGraph checkpoints”

Should not

“A one-shot classifier”