Running OpenClaw on Windows in gateway dispatch mode (warm gateway,
no --local). 4 registered agents: 1 chat supervisor + 3 workers
(filings, financial-statements, intrinsic-value). Workers
exec deterministic Python runners; we dispatch via
openclaw agent --agent <slug> --session-id <fresh uuid> --message <json> --json --timeout 1800.
Pipeline currently serial โ one company through 8 stages, one
stage at a time. We want to dispatch 5-20 concurrent runs against
the same agent identities, each with its own session UUID.
Questions:
- Can one agent identity handle N concurrent dispatches with
distinct session IDs, or does the gateway serialize per-agent? - Recommended shape: one agent identity + N concurrent sessions,
or N copies of the same agent with 1 session each? - What signals indicate the ceiling โ 5xx, silent fallback to
meta.transport=embedded, stderr patterns? - Any tuning knobs before we push past ~5 concurrent?