escapeboy/agent-fleet-o/fleetq-run-experiment

Run a FleetQ experiment

An experiment is FleetQ's unit of autonomous work: it moves through a 20-state pipeline (scoring → planning → building → approval → executing → metrics → evaluation) with budget enforcement and an audit trail at every step.

Requires a working connection — see the fleetq-connect skill first.

Create and start

MCP tools:

REST equivalent:

POST {{APP_URL}}/api/v1/experiments
POST {{APP_URL}}/api/v1/experiments/{id}/transition
GET  {{APP_URL}}/api/v1/experiments/{id}

Monitor progress

• experiment_steps — per-step execution detail.
• experiment_stage_telemetry — timing and stage-level metrics.
• experiment_cost — credits consumed so far.
• experiment_valid_transitions — which transitions are legal right now.

Call this before attempting a transition; the state machine rejects anything not in the map.

Intervene

Rules that will bite you

• Transitions are validated. ExperimentTransitionMap rejects illegal

moves — always check experiment_valid_transitions rather than guessing.

• Terminal states are final: completed, killed, discarded,

expired. There is no transition out; create a new experiment.

• Budget is enforced before execution. An experiment pauses

automatically when the team's credit balance cannot cover the reservation (reservations use a 1.5× multiplier). Check budget_summary first if you expect long runs.

• Approval gates block execution. If the experiment enters

awaiting_approval, a human (or an agent with approval permission) must act via approval_approve before it proceeds.