Create a Launch queue at wandb.ai/launch, choosing an entity, queue name, and target resource (Docker, Kubernetes, SageMaker, or Vertex AI)
Start a Launch agent that polls the queue: `wandb launch-agent --queue <queue-name> --entity <entity-name>` (or deploy the agent via the official Helm chart, setting `agentConfig`); for git/code-based jobs, set `builder.type` to `docker` or `kaniko` and a registry `destination` in the agent's `launch-config.yaml`, since the default builder is `noop` (no build)
Submit a job to the queue from the W&B UI Jobs tab (Launch drawer -> select queue) or directly via CLI with `wandb launch --uri <...> --queue <queue-name>`
Monitor execution from the W&B UI or the polling agent's terminal output as it builds/pulls the image and runs it on the target resource
Known gotchas
Create the queue before starting the agent — the agent needs a queue with a compatible target-resource configuration to have anything to poll
Image-based jobs need no build step, but git-based and code-based (artifact) jobs require the agent to be configured with a builder (Docker or Kaniko); the Helm chart's default `builder.type` is `noop`, which will not build anything
Every Launch job must call `wandb.init()` to complete successfully and register a tracked run in a W&B workspace, even for non-training tasks like batch inference or deployment jobs
Give your agent this knowledge — and 15,500+ more routes
One MCP install gives any agent live access to the full route map across 5,700+ domains, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp
Need this verified for your stack — or a route we don't have yet?