How AI is changing work inside Anthropic

What happens when the people building cutting-edge AI turn the tools on themselves? Instead of debating hypotheticals, Anthropic just surveyed 132 of its own engineers and researchers to see how Claude is transforming their daily work.

The results show massive productivity leaps, unexpected workflow shifts, and a surprisingly honest look at developer anxiety.

Massive output gains, but zero autopilot

Anthropic employees report using Claude in 59% of their daily work, driving an average 50% productivity boost. Both numbers have more than doubled year-over-year. Following the internal rollout of Claude Code, merged pull requests per engineer jumped by 67%.

Interestingly, AI isn't just speeding up existing backlogs; it is expanding the surface area of work. Roughly 27% of Claude-assisted work consists of tasks that employees simply would not have tackled otherwise, such as building disposable internal dashboards, running speculative experiments, or cleaning up technical debt. Fixing minor workflow papercuts alone accounted for 8.6% of Claude Code tasks.

The agent is also getting longer leashes. Six months ago, Claude Code could handle roughly 10 autonomous actions before requiring human steer. Today, it routinely clears 20 actions. High-level planning tasks assisted by Claude surged from 1% to 10%, while feature implementation jumped from 14% to 37%.

Yet despite these gains, true full delegation remains virtually nonexistent. Most engineers reported they can fully delegate only 0% to 20% of their tasks without verification. Claude is treated as a tireless junior collaborator that still requires relentless oversight.

The atrophy trap and fading mentorship

Anthropic engineers describe becoming instantly full-stack. Alignment researchers now spin up front-end visualizers, and security teams quickly parse unfamiliar database repos.

However, this newfound breadth brings real friction. Multiple employees voiced concerns over skill atrophy. When generating complex code takes seconds, engineers worry they are losing the muscle memory needed to understand deep system mechanics, which could eventually compromise their ability to properly supervise the AI.

Social dynamics inside the office are also shifting. Claude has largely replaced colleagues as the default first stop for technical questions. Senior staff noted that junior developers rarely drop by with questions anymore, threatening the informal apprenticeship model that has sustained software engineering for decades. As one researcher put it, while short-term optimism is high, there is a lingering fear that AI will eventually handle everything, rendering developers obsolete.

Why it matters

This study serves as a much-needed reality check against executive fantasies of fully autonomous engineering teams. Even at Anthropic, with direct access to Claude Opus 4 and internal tooling, engineers only fully trust AI with a sliver of unreviewed tasks. The immediate value for engineering teams is clear: use agents to crush boilerplate, build internal tools that were previously too expensive, and accelerate debugging, but do not drop the review standards.

The industry-wide challenge lies in developer education. Software engineering has always relied on junior developers struggling through low-stakes bugs to build the intuition required for senior architecture. If Claude takes over all introductory tasks and junior engineers stop asking questions, the talent pipeline breaks.

We think Anthropic's transparency here is telling. It confirms that the AI coding revolution is less about immediate layoffs and more about a painful skill transition, where typing syntax becomes cheap, but verification taste becomes the ultimate moat.