Steps of AI Adoption

A maturity ladder for AI-assisted software engineering, summarized from a 2026 framework by Boris Cherny, the creator of Claude Code (original artifact).

The idea: teams climb five steps, each multiplying how many AI agents one engineer directs — and at every step the bottleneck moves somewhere new.

Step	Your role	Agents at once
0 · Gated	—	0
1 · Assisted	Pair partner	~1
2 · Parallel	Orchestrator	~10
3 · Supervised autonomy	Manager of managers	~100
4 · AI-native	Steering by intent	~1,000+

The steps in brief

• 0 · Gated — approvals, gateways and governance keep real AI tooling out of engineers' hands, and whatever does get produced has no official place to live. Escaping this step is less about technology than executive alignment.
• 1 · Assisted — one engineer pairs with one supervised agent and reviews everything it does. Fast, but synchronous: your attention is the ceiling. The way up is a self-verification loop you actually trust (tests, build, lint, end-to-end checks) plus automated code review.
• 2 · Parallel — one engineer runs a handful of isolated agents at once and reviews final diffs instead of keystrokes. The bottleneck becomes your own review and steering bandwidth. The way up: feed agents richer context, and let agents kick off agents.
• 3 · Supervised autonomy — agents write nearly all the code, and proactive background loops handle maintenance nobody scheduled. The question shifts from "did you read the code?" to "what context was the model missing?". Trust in the loop and the team's decision throughput now gate progress.
• 4 · AI-native — the loop closes: most agents are launched by other agents, hundreds to thousands run at once, and humans steer by intent and monitor by exception. The remaining craft is choosing what to automate and matching guardrails to each kind of work.

Two threads run through the whole ladder: build verification you trust before scaling autonomy, and grow guardrails (spend caps, code and security review, sandboxing) in step with agent count.