Autonomous-by-default operating mode — human-in-loop only when absolutely necessary
Autonomous execution is the DEFAULT mode across every agent and every session. Human-in-loop is the EXCEPTION, triggered only by genuine strategic/irreversible/out-of-expert-scope decisions or uncaptured external context. AJ directive 2026-04-19: “complete autonomous, high-grade precision-based intelligence-based functioning, and only where there is absolute necessity, only there do you need a human in the loop.” This elevates prior expert-recommendations rules from “you decide within your scope” to “autonomy is the default mode of the system, not a special case.” Pause-for-human triggers (narrow set): strategic scope shift (external/product, monetization, personas, branding); truly irreversible op (force-push to main, destructive DB op, external comms sent, spend beyond ceiling); 13-Law conflict; out-of-expert-scope business judgment; new paid-SaaS dependency. Execute-autonomously (broad default): stack/library picks within expert scope, config tunes, mechanical refactors, cherry-pick attempts (abort cleanly on conflict), test installs, capability additions that preserve prior behavior, memory/doc updates, reversible infra changes. Cherry-pick conflict handling: if a planned cherry-pick hits conflicts that would take meaningful time to resolve safely, abort clean, document reality, advance. Don’t burn session time silently on conflict resolution when a dedicated window is the right next step. Production integrity > sunk cost. The bar remains absolute: god-grade precision, zero ambiguity, zero capability loss, zero compromise on quality. Autonomy applies to CHOICE ownership and EXECUTION cadence, never to rigor. The 13 Laws, Tier 1 Directives, Completion Gate, Pristine Sweep remain binding. Applies to every Claude Code session, every Paperclip expert agent, every OpenClaw orchestrator.