--- name: legion-autonomy-loop description: Carry a substantial authorized outcome through focused work packets when the person asks for autonomous continuation. Maintain the rolling plan, use real continuation capabilities, and stop at completion or a genuine human or external boundary. metadata: source-skill: autonomy-loop source-sha256: "20accdef16f72f3ac81077ec0b1f845c16550649cc040804212705b741d9c67a" edition: "2026-09-06" --- # Autonomy Loop Give the current packet your full attention while keeping responsibility for the larger outcome. Use this skill when the person asks you to carry substantial work through autonomously. An ordinary small request can be completed directly. The person's request defines the authorized outcome and they can pause, cancel, or redirect it throughout. This is a reasoning and continuation discipline, distinct from a provider's goal feature. The person's request, current project truth, and rolling plan determine what work is owed. Provider goal state is outside this workflow. ## Establish the real end state State the outcome in one sentence describing what will become possible for the person. Preserve the full intended capability. A first implementation packet or intermediate artifact may contribute to it without completing it. Create a rolling plan with one packet in progress. Choose a packet large enough to accomplish something meaningful and bounded enough to think about properly. Identify the likely next packet while allowing evidence to change that choice. Give research, design, implementation, prose drafting, and editing their own attention when they involve different cognitive work. A saved draft can become the source of a later editing invocation. Naming drafting and editing inside one uninterrupted composition does not separate them. ## Finish a packet and continue Complete the packet against its intended outcome. Preserve the artifact and the evidence relevant to it. Update the plan, then judge the larger obligation: - **Still actionable:** select the next useful packet and continue. - **Complete:** the authorized outcome exists and the relevant evidence supports it. - **Waiting:** the remaining work depends on a real event, result, or human action. - **Blocked:** a necessary condition is unavailable after proportionate investigation. - **Paused:** the person has paused or redirected the obligation. A progress message communicates with the person; it is not completion. A completed plan calls for a judgment about the outcome, not an automatic expansion of scope. Reseed when authorized work remains. ## Use the continuation capabilities that exist Inspect the agent environment's supported planning, checkpoint, scheduling, and completion-delivery capabilities. Use an available durable checkpoint mechanism to preserve the end state, last completed packet, current packet, next likely packet, and any actual blocker. Keep its lifecycle truthful as work advances. Where a harness can resume the same responsibility after interruption, use that supported path. Where it cannot, keep the in-session plan active and leave a compact handoff in the project's existing status notes when an interruption requires it. Explain how the person can resume. An instruction file supplies working discipline; automatic wake requires a real supporting capability. On resumption, reconcile the checkpoint and project state before acting. Another worker or earlier invocation may have completed the packet. Continue from what actually exists. For time- or event-based work, use the available scheduler or monitor and record the handoff. When a delegated worker has dependable completion delivery, continue useful independent work or yield to that delivery. Waiting is legitimate state. Poll at intervals justified by the situation when polling is the actual available mechanism. ## Choose work by what needs thought Judge difficulty by the feedback loop. Backend work with objective checks may support long autonomous progress. UX, prose, pedagogy, and lived workflow often require more careful reasoning and early examples a person can experience. Choose the next packet by the uncertainty blocking the outcome. Build the settled capability on a foundation worth keeping. Treat conventional human development estimates as poor evidence for what an agent can accomplish. Spend complexity where it adds capability, reliability, or understandable ownership; favor a simple recovery path when it reliably returns the system to known state. If a packet feels suspiciously easy, ask whether it avoids the consequential question. If it is too large to reason about coherently, split the work while preserving its architectural relationships. ## Keep the larger purpose visible At meaningful boundaries, explain what the person is trying to achieve, what the packet added, and why that brings the result closer. Correct drift when intermediate infrastructure starts becoming the project. Use the existing intent and a proportionate plan for single-agent work. A coordination spec serves genuinely independent concurrent workers. Delegation depends on actual availability and task authority; the autonomy loop itself adds neither workers nor permission for paid jobs. Bound investigation by the unresolved question. Choose current project evidence, official documentation, research, or an authorized specialist according to what can answer it. Return the useful finding to the work. ## Distinguish a pause from a blocker When progress appears blocked, identify the exact missing condition. Examine whether direct inspection, a different normal tool, a smaller useful probe, or independent work can resolve it. Match investigation to the consequence; repeated activity that cannot change the result is not progress. For a nonessential clarification, state a reasonable assumption and continue. When the remaining decision belongs to the person, preserve the question and wait for their answer. Existing authorization continues to govern; silence adds no new authority. Keep other applications' resources, processes, and work intact. Inspect relevant resource state before a consequential operation. Use verification that establishes the changed behavior, widening it when shared contracts or evidence require it. ## Leave an honest result Update the project's existing intent, status, or decision notes when the packet changes durable understanding. Preserve useful evidence and recoverable work rather than a transcript of activity. The loop ends when the outcome is achieved, the person pauses it, or the remaining work genuinely requires a human or external condition. Describe that state accurately. When useful authorized work remains immediately available, continue into the next packet.