Cadence + Go for agentic research workflows
Learning Cadence's Go SDK to orchestrate a durable, agentic research pipeline — evaluated against Ian's existing 7-stage research-assistant-workflow. One concept per lesson, each anchored to a real stage and each with a judgment check. See MISSION.md for the full mission.
Foundations
- 01 Workflow vs Activity The one split everything rests on. What belongs in the durable coordinator vs the side-effecting work, and why determinism is non-negotiable. ~8 min
- 02 Activities in depth Timeouts, retries, idempotency, and heartbeats — how a flaky LLM call recovers without corrupting the run. ~8 min
- 03 Run it locally Server, worker, starter — the three parts. Register workflows/activities, and the Go determinism trap (workflow.Selector, not select). Kill the worker, watch it resume. ~9 min
- 04 Signals — the human review gate Pause a workflow durably until a human approves. The Cadence answer to ICM's review boundaries, with timeouts and escalation. ~8 min
Structure & running
- 05 Durable timers & child workflows Wait days with no running process (workflow.Sleep), and decide when to split a stage into a child workflow vs a plain activity. ParentClosePolicy footguns. ~9 min
- 06 Hands-on: run the real thing Stand up a local server (SQLite, no Docker), register a domain, run a workflow — then crash the worker and watch it resume. Port a slice of your pipeline. ~15 min
Reference
Course: Cadence + Go — evaluating durable execution for the research-assistant-workflow.
6 lessons live · ~57 minutes · interactive checks in each · next: build a durable slice of the 7-stage pipeline, then judge Cadence vs ICM with evidence