Boris Cherny outlined what he described as a four-step path to AI adoption, arguing that teams can move forward only by "finding and breaking down the next set of bottlenecks" and adding the next layer of guardrails, rather than relying on more tokens alone.
In the thread, Cherny mentioned a stack of features and workflows tied to that progression, including automated code review, security review, multi-agent interfaces, /loop, /batch, dynamic workflows and worktree isolation for subagents. He also pointed to ways for Claude to verify its own work end to end, while stressing that the goal is to automate entire classes of work in a way teams can trust.
Cherny also challenged simple usage metrics as a way to measure AI adoption. Usage dashboards, he wrote, track activity, not return. A better test, according to him, is whether the engineering effort saved by the system is work that would otherwise have been done manually, and at what cost in engineering hours.
Near the end of the thread, Cherny wrote that Anthropic is "on step 3 and pushing toward 4," adding that he had "just hit level 4" personally. The thread drew a range of responses from people describing themselves as stuck in stage 1, moving from 2 to 3, or blocked by permissions, integrations and cost. A few replies also raised questions about harness infrastructure and how the step-3-to-step-4 transition works in larger teams.
Source: X
