Steps of AI Adoption: Boris Cherny's 5-Stage Model

Key takeaways

1. Core idea: you don't move up a level by spending more tokens. Each tier has its own bottleneck; moving to the next requires breaking it AND building the next set of guardrails that makes the output trustworthy. « It's not about a single feature, but rather using the right features with the right guardrails. »
2. The model's backbone = the order of magnitude of agents driven, and the role it imposes. 0 → ~1 → ~10 → ~100 → ~1,000+ agents, i.e. blocked engineer → pair programmer → orchestrator → manager of managers (org tree) → VP steering by intent. That's the table's true "essence": an exponential rise in parallelism that transforms the job, not just velocity.
3. Stage 0 — Gated (0 agents). Only older/lightweight approved models, latency piling up through gateways + custom auth, no MCP governance, locked-down/process-heavy access, no infrastructure to host code produced by Claude (outputs stay local). Bottleneck: legacy security/approvals, an obsession with cost-per-token at the expense of outcomes, no real technical voices in the decision. Product: Claude.ai chat. Guardrails: SSO/SCIM + RBAC, org budget caps, deployment within existing approvals/IAM, data governance. 0→1: executive/buyer alignment, escalating blockers, frameworks to launch Claude safely.
4. Stage 1 — Assisted / "you + one agent" (~1). One engineer, one agent, almost everything supervised (fast pair programmer), one session at a time, reviewing nearly every change before merge. Unlock: what used to fill an afternoon now wraps up between two meetings. Bottleneck: your attention — low trust + no self-verification → you read everything, synchronous work (you watch Claude instead of moving on). Products: Claude Code (Desktop/CLI/IDE), Cowork, Design, usage via API/Bedrock/Vertex/Microsoft Foundry, dashboard + Analytics API, Compliance API, Plan mode. Guardrails: per-seat spend caps, centrally managed model/effort and policy, OpenTelemetry export to the SIEM.
5. 1→2 (key transition):. running several agents at once; a trustworthy self-verification loop (tests + build + lint + e2e on a real dev environment); Auto mode to stop being blocked by permission prompts; automating code review.
6. Stage 2 — Parallel / Orchestrator (~10). An engineer orchestrates 5–10 agents in parallel, each on its own worktree/git checkout. Claude verifies its own work (tests, build, lint, security scan) before you see it; Auto mode always on; code review + security review by default. You review final diffs, not keystrokes; the maintenance backlog shrinks; Claude writes most of the code. Unlock: a multi-week backlog becomes an engineer's afternoon of orchestration. Bottleneck: reviewing (checking six parallel streams) + juggling prompting/steering. Products: Auto mode, Agent view, Claude Code Review, Claude Security Review, Claude Code Mobile + Desktop cloud execution, Claude Tag (an isolated task), worktree isolation CLI/Desktop, Remote control (steering from your phone). Guardrails: team usage Analytics, quality enforcement (lint/tests/typecheck), Claude-powered e2e verification (Chrome extension or iOS/Android simulator MCP), the same quality bar for human code and agent code, pre-approving safe bash/MCP in settings.json.
7. 2→3:. giving Claude the means to retrieve context (reading code, wikis, discussions); agency + code review speed (agents can touch other teams' code); breaking work down into loops and routines; letting Claude launch Claude.
8. Stage 3 — Supervised autonomy / Manager of managers (~100). Claude writes all or almost all of the code; the question shifts from "did you read the code?" to "what context was the model missing, and how do we fix that for next time?". Unlock: Claude proactively does what used to be launched by hand; maintenance/cleanup runs continuously in the background. Bottleneck: trust in the loop and the team's decision throughput — the agent tree is too deep to be babysat; the trap is scaling the number of agents before the loop has earned trust. Token efficiency to monitor (OTel/Analytics), a culture of experimentation + cost control once internal usage hits PMF. Test: "is this something an engineer would have done?". Products: subagents with worktree isolation (no collisions), Routines, /loop, /batch, /goal, dynamic workflows, Claude Tag (watches a channel/source and launches tasks). Guardrails: automated code review + security review, agent sandboxing, CLAUDE.md + Skills to encode standards, tuning the Auto mode classifier, managing token use (model selection, advisors, LSP, splitting CLAUDE.md into lazy Skills).
9. 3→4:. automation at scale for domain-specific use cases (code migration, fuzzing, feature building, feedback remediation).
10. Stage 4 — AI-native / VP steering by intent (~1,000+). A fully closed loop; most agents are launched by Claude; hundreds to thousands of agents running; steering by intent and monitoring by exception. Unlock: a quarter-long migration becomes a workflow you launch and check on. Bottleneck: identifying and automating work at scale + enforcing the right guardrail per type of work. Products: Claude Agent SDK (building/scheduling agents programmatically), Claude Tag (present in most Slack channels, replies to posts). Guardrails: cost controls for automation, model selection for automation.
11. Anti-vanity metric (steering). Usage (dashboard) measures activity, not return. Better question: would we have invested engineering effort on this anyway, and if so, how many manual engineer-hours would it have cost? → that's the ROI. The jackpot: when fixing/maintaining moves to the background, teams do "things that weren't even within the realm of possibility before".
12. Self-placement benchmark. Anthropic is at stage 3, pushing toward 4; Boris Cherny states he's at level 4. Useful as an honest benchmark: even the vendor isn't "at the very top".
13. Related. companion piece to [[fiches/2026-07/cherny-wu-reflecting-year-claude-code-2026-07-17\|Cherny & Wu — « Reflecting on a year of Claude Code »]] (same author, shipping becomes cheap → value shifts to judgment); "agent manager / orchestration" corpus (Sankar's great flattening, Martignole's token manifesto); guardrails & verification (compound engineering, self-verification loop).