Skip to content
Ankole

Workers

An Agent Computer Worker is the computer where Agents do their work. The control plane manages and assigns work. A Worker runs Agents, tools, and file operations. One deployment instance can connect to one or more Workers.

View Workers

Open Console → Workers. Each row shows:

  • State: whether the Worker can accept new work;
  • Slots: the largest number of Agent turns that it can run at the same time;
  • Active turns: turns that are running now;
  • Last heartbeat: the last time that the control plane received Worker state;
  • Version: the Ankole version on the Worker.

The control plane can assign work while at least one Worker is ready. A Worker with no recent heartbeat usually has a stopped container, a network problem, or an authentication failure between the Worker and control plane.

Select Browse files to inspect Agent files on that Worker. See File management for the procedure.

If no Worker is ready

Check these items in order:

  1. The Worker container or Pod is running.
  2. Worker logs do not show a control-plane address, authentication, or connection error.
  3. The control plane and Worker use the same ANKOLE_RUNTIME_FABRIC_WORKER_AUTH_KEY.
  4. The Worker can reach the control plane Runtime Fabric address.
  5. The control plane and Worker use compatible Ankole versions.

For Docker Compose:

docker compose ps agent-computer-worker
docker compose logs agent-computer-worker

For Kubernetes:

kubectl -n ankole get pods
kubectl -n ankole logs deployment/ankole-agent-computer-worker

The exact resource name depends on the release name. If the command cannot find the object, use kubectl -n ankole get deployments to find it.

If Workers are ready but work stays queued

Compare Active turns with Slots. If every Worker uses all of its slots, new work waits for capacity.

A short wait is normal. For a persistent queue, first confirm that the LLM Provider is not slow. Then decide whether to add Workers, increase per-Worker capacity, or reduce concurrent work. See Performance tuning.

Restart a Worker

Restart a Worker only when it does not respond and its logs show that it cannot recover.

Docker Compose:

docker compose restart agent-computer-worker

Kubernetes:

kubectl -n ankole rollout restart deployment/ankole-agent-computer-worker

A restart interrupts turns that run on that Worker. The control plane handles them according to their retry policy. After the restart, check the result under Console → Conversations or Background Agent Jobs.