Autonomous mode (/autonomous) continues within configured turn, token, and time budgets and can run user-defined quality gates; a passed gate checks only what that gate verifies, and reaching a budget limit does not imply task success.[1]
In v0.1.5, goals were moved into a bundled goal Python skill backed by a typed host bridge, leaving IPython as the only built-in tool.[2] Autonomous mode was introduced in v0.3.0 with host-side continuations, configurable limits, and quality gates for evaluator-controlled runs.[3] A --goal seeding bug was fixed in v0.5.0: CLI --goal sessions now show the objective to the model, which had previously been invisible to first turns and continuations because the goal-context message was never injected.[4]
HarnessOptions exposes an autonomous field of type AgentAutonomousConfig and an autoRefineReviewer field, both passed to AgentSession to enable autonomous mode with quality gates in tests — see Test harness for full harness API coverage.[5]
Sources