What problem does it solve? Running autonomous and parallel agent work in Sonder involves layered systems—autopilot runs, master orchestrator fleets, model fanout, and a multi-PC Ollama worker pool—each with its own lifecycle states, capacity limits, cancellation semantics, and failure modes. This Skill provides the verified operational knowledge to start, monitor, cancel, retry, and troubleshoot all of them without guessing at internal behavior. ## Core Features & Use Cases - Autopilot run management: Start, pause, resume, and cancel restart-safe autonomous goal runs with hard bounds on cycles, checkpoints, and task output, including crash recovery via explicit resume of interrupted runs. - Fleet orchestration and capacity planning: Delegate parallel subagents through master_orchestrate with computed worker slots based on CPU, RAM, GPU VRAM, and Ollama batch width, plus cooperative cancellation and idempotent retries through the durable fleet ledger. - Protected objective runs and inference pooling: Enforce provenance contracts requiring tool evidence for fleet claims, run model fanout across local and cloud models, and configure a multi-PC Ollama worker pool with TLS validation, circuit breakers, and pre-response failover. - Use Case: A user says "use 24 workers to audit this repository in parallel." The Skill guides checking master_capacity for the real worker slot bound, launching a fleet-mode run confined to the project root, monitoring with master_status, and cooperatively cancelling or retrying agents as needed. ## Quick Start Ask the assistant to check fleet capacity with master_capacity and then start a delegated fleet run confined to your project root using master_orchestrate.
