Addy Osmani proposes a six-level autonomy framework for agentic engineering, separating two dimensions: single-agent agency (how far an agent operates independently) and multi-agent orchestration (coordinating many agents in parallel). The six levels range from Level 0 (assist/autocomplete) through Level 5 (managed-by-exception orchestration with manager agents dispatching worker fleets). Each level is defined with verification requirements, failure modes, and appropriate use cases. The framework critiques single-axis ladders (like Steve Yegge's) as insufficient for multi-agent work. Key guidance includes defining agent contracts upfront (goal, scope, stopping conditions, evidence, budget), tracking metrics like mean time between interventions and defect escape rate, and avoiding four anti-patterns: autonomy as status, permission laundering, summary substitution, and fleet cosplay. The core thesis is that high autonomy is not about removing humans from the loop but shifting them from executing every step to directing overall goals, with verification always being the bottleneck.
