Addy Osmani proposes a two-axis framework for understanding AI agent autonomy in software engineering: agency (how far a single agent operates independently) and orchestration (how many agents are coordinated simultaneously). The framework defines six levels from Level 0 (assist/autocomplete) through Level 5 (managed-by-exception orchestration with agent factories). Key insights include: high autonomy is about shifting humans from doing every step to deciding direction; every agent run should be preceded by a contract defining goal, scope, non-goals, tools, stopping conditions, evidence, escalation, and budget. The post identifies four anti-patterns (autonomy as status, permission laundering, summary substitution, fleet cosplay) and recommends specific metrics to track per autonomy level. References Anthropic research showing humans make ~70% of planning decisions while Claude handles ~80% of execution across ~400K sessions.
Table of contents
The climb: three eras and a single stackThe six levels in detailRisk and reversibility set the ceiling.Metrics make autonomy just a little more reliableThink about readinessFour anti-patterns6.2K Impressions1 Comment