Multi-Agent Orchestration

Hosted orchestration on nfltr.xyz runs explicit task-mode phases such as implement, review, summarize, and integrate via orch spec decompose and orch start spec.

Decompose → run

$ nfltr orch spec decompose --goal "Ship docs track and fix flaky tests in parallel slices"

$ nfltr orch spec decompose --goal "Ship docs and tests" | nfltr orch start spec - --watch

$ nfltr orch start spec ./my-objective.json --watch

Phases

Each phase declares task_mode or execution_profile.mode. Some modes mutate artifacts, some review or verify them, and others summarize or integrate child outputs. Trivial goals may be a single phase.

Parallel slices

$ nfltr orch start batch ./two-slices.json --watch

Use artifact_request / handoff_request for large files. Ask for fulfill_worker_inventory_request when available_workers is empty.

$ nfltr orch fleet workers --json

$ nfltr orch task result --task task-review-slice-a

See Dashboard walkthrough and Troubleshooting.

Practical tips

NFLTR persists task history on the hosted control plane at nfltr.xyz, so reconnecting planners and dashboard viewers see the same timeline. Share runs with teammates using task ids and dashboard deep links instead of pasting full JSON envelopes. For production-impacting objectives, configure approval gates so a human confirms in the Atlas attention rail / task drawer before work continues. Combine the CLI examples above with the Causal Operations Atlas (causal canvas, Trace Lab) when you need mobile monitoring or inner-LLM turn detail documented in the observe orchestration guide.