01 · Comparison
Five working levels from manual implementation to bounded automation. The useful question is not how autonomous the tool looks; it is whether review, recovery, and ownership remain clear.
5 working levels · first-person editorial guidance
02 · Working spectrum
Level 01
A person performs the implementation directly.
What AI does
No required role
What I do
Performs and reviews the complete change
A person owns architecture, implementation, verification, and recovery directly. · Tools — Editor, documentation, test tools
Level 02
AI suggests small changes while a person drives each decision.
What AI does
Suggests local changes
What I do
Chooses, reviews, integrates, and verifies
AI proposes a small edit or explanation while a person controls the sequence and accepts each material change. · Tools — Inline suggestion or chat assistant
Level 03
AI drafts a bounded change for human review and integration.
What AI does
Drafts one bounded change
What I do
Defines the boundary and verifies the result
AI drafts a bounded function, file, or task from an explicit request. A person reviews the result and owns integration. · Tools — Context-aware coding assistant
Level 04
An agent executes a scoped plan inside explicit verification gates.
What AI does
Plans and executes inside named gates
What I do
Owns architecture, permissions, review, and release
An agent follows a scoped plan across multiple files and reports verification evidence before a person accepts the change. · Tools — Repository-aware coding agent
Level 05
A repeat workflow runs with observability, limits, and a recovery owner.
What AI does
Runs a known repeat process
What I do
Owns policy, exceptions, monitoring, and recovery
A repeat workflow can run without a person driving every step because limits, telemetry, review thresholds, and recovery ownership are already defined. · Tools — Workflow orchestration and monitoring
Answer from current operating evidence, not tool capability.
My recommendation · reviewed Jul 13, 2026
Move up only when the next level keeps the work more reviewable, not merely faster. Level 3 is often right for a new or weakly tested codebase. Level 4 needs a bounded task, useful verification, and a person who can reject the result. Level 5 belongs only to repeat work with operational limits and recovery ownership.