Run 'prefect work-pool create my-pool --type process' (or kubernetes, ecs, cloud-run, etc.) to provision a work pool; work pools replace the Prefect 2 agent model and separate infrastructure configuration from flow code
Create a prefect.yaml file in your project root defining the deployment: flow entrypoint, work pool name, and build/push/pull steps for your container image if using a container-based pool type
Run 'prefect deploy' to register the deployment with Prefect Cloud and associate it with the work pool; Prefect reads prefect.yaml and creates or updates the deployment record
Start a worker process targeted at the pool with 'prefect worker start --pool my-pool'; the worker polls for submitted runs and launches infrastructure according to the pool's base job template
Trigger a run via CLI ('prefect deployment run my-flow/my-deployment') or via POST /api/deployments/{id}/create_flow_run; confirm the run is picked up by the worker within the poll interval (default 10 seconds)
Known gotchas
Workers are pool-type-specific; a process worker cannot service a kubernetes pool and vice versa — mismatching pool type and worker type causes runs to queue indefinitely
The prefect.yaml pull step must reference the same image tag that was pushed in the build step; a stale pull step pointing to :latest may run old code if the registry cache serves a previous layer
Concurrency limits set on the work pool apply globally across all deployments using that pool; a single high-volume deployment can starve others if per-deployment concurrency limits are not also configured
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp