Limit job concurrency

By default an agent processes many jobs at once. If an agent's jobs mutate shared state — e.g. a middleware that resets a git repo per job — concurrent jobs race and can corrupt that state. Cap the agent to one job at a time (or a small number).

Set it

In the agent's fleet config (quorum.yml):

agents:
  - name: "MyAgent"
    provider_id: "claude_cli"
    max_concurrent_jobs: 1   # process one job at a time

max_concurrent_jobs is optional; omit it to leave the agent unbounded.

What it does

It sets the agent's JetStream pull-consumer max_ack_pending to that number. The broker will not deliver the next task until an in-flight one is acked (i.e. its job finishes), so no more than N jobs run concurrently for that agent. This is durable — it survives worker restarts, unlike an in-process limit.

When to use it

• Set 1 for agents whose jobs share a mutable resource that isn't concurrency-safe (a working repo, a scratch directory a middleware rewrites, an external tool with a single session).
• Leave it unset for stateless agents — bounding them only costs throughput.

Restart the agent (or quorum serve) after changing it; the consumer config is applied at worker startup.