What did Anthropic reveal about how its engineers use Claude Code?

Anthropic published a July 2025 report of team case studies, a December 2025 study of 132 engineers with about 200,000 Claude Code transcripts, and an October 2026 essay by Felix Rieseberg. They show engineers handing the agent larger tasks with fewer interventions, non-engineers building tools, and verification against real systems as the habit that makes autonomy safe.

How much more productive are Anthropic engineers with Claude Code?

In Anthropic's August 2025 survey, self-reported productivity gains rose from about 20 percent to about 50 percent year over year, and 14 percent of respondents reported gains above 100 percent. Anthropic stresses these are self-reports with sampling and recall bias, and that much of the gain appears as additional output rather than fewer hours.

Do non-engineering teams at Anthropic use Claude Code?

Yes. The July 2025 report describes growth marketing building an agentic workflow that analyzes ad performance and generates variations, the legal team prototyping a phone tree to route staff to lawyers, and product design turning Figma designs into code. Data scientists without TypeScript experience built React visualizations for model performance.

What risks did Anthropic's internal study find?

Engineers worried about skill atrophy as they delegate more, a supervision paradox where reviewing the agent requires the skills that fade with disuse, and less collaboration, since Claude became the first stop for questions once asked of colleagues. Some also described uncertainty about how the profession will change as work shifts toward directing agents.

Can a smaller team copy Anthropic's Claude Code workflow?

The habits transfer, the numbers may not. Start with maintained instruction files, give the agent tests or real systems to verify against, ask for outcomes rather than steps, and review its work with an outside check. Anthropic's results come from engineers with early access to frontier models, so measure your own baseline before and after.

How Anthropic's Own Teams Use Claude Code, According to Anthropic

Anthropic has published three accounts: a July 2025 report on how its teams use Claude Code, a December 2025 study of 132 engineers and about 200,000 transcripts, and an October 2026 essay by Felix Rieseberg. Together they show longer autonomous runs, fewer human turns, more new work rather than less time, and verification as the enabling habit.

The three sources, and what kind of evidence each is

Most videos about how Anthropic engineers work are summarizing one of three documents Anthropic has put out itself. They are different kinds of evidence, and they should be weighed differently.

How Anthropic teams use Claude Code (July 24, 2025) is a set of team case studies: infrastructure, product engineering, security, inference, data science, growth marketing, legal and others. It is anecdotal by design. It tells you what people tried, not how often it worked.

How AI is transforming work at Anthropic (December 2, 2025) is the closest thing to a study. Anthropic surveyed 132 engineers and researchers in August 2025, ran 53 interviews, and analyzed about 200,000 internal Claude Code transcripts from February and August 2025.

How to be more ambitious: a tutorial for humans with models (October 13, 2026) is a personal essay by Felix Rieseberg, who leads engineering for Claude Cowork and Claude Code Desktop at Anthropic.ai, Cowork and Claude Code Desktop. It is opinion, informed by practice.

What the usage data actually shows

The transcript analysis is the part worth reading closely, because it is measured rather than recalled. Between February and August 2025, Anthropic reports that the maximum number of consecutive tool calls per transcript rose 116 percent, from 9.8 to 21.2. Human turns per transcript fell 33 percent, from 6.2 to 4.1. Average task complexity, on a 1 to 5 scale, rose from 3.2 to 3.8. Feature implementation grew from 14.3 to 36.9 percent of tasks, and design or planning from 1.0 to 9.9 percent.

Read together, that is a shift in the contract between engineer and agent. Six months earlier the agent was being steered every few steps on smaller tasks. By August it was being handed larger, harder tasks and left to run longer before a person stepped back in. The model improved over that period, so not all of this is habit. But the habit has to change for the capability to be used: an engineer who still intervenes every two tool calls gets none of it.

The output effect

The survey adds the finding most rollout plans miss. Respondents said 27 percent of their Claude-assisted work consisted of tasks that would not otherwise have been done, such as nice-to-have dashboards and scaling projects. Engineers also reported that assisted work increased output volume more than it reduced time spent. If you measure a rollout only by hours saved, you will undercount what changed.

What the teams did that transfers

The team report is most useful as a catalogue of patterns, several of which work in any organization.

Onboarding through instruction files. Anthropic's infrastructure team describes new data scientists asking Claude Code to read the repository's CLAUDE.md files and explain pipeline dependencies, which the report says replaced data catalog tools for that purpose. The transferable point is that a maintained instruction file doubles as onboarding documentation.

Screenshots as context during incidents. The data infrastructure team fed dashboard screenshots to Claude Code during a Kubernetes outage. It walked them through the Google Cloud console until they found pod IP address exhaustion, then supplied commands to add an IP pool.

Specialists crossing into unfamiliar code. The inference team used it to translate tests into Rust and explain model-specific functions. Data scientists without TypeScript built React visualizations.

Non-engineers building tools. Growth marketing built a workflow that processes CSV files of ads, flags underperformers and generates variations with two subagents. The legal team prototyped phone tree systems to route staff to the right lawyer.

The essay's argument: ask for outcomes, then verify

Rieseberg's essay is about the gap between people who get remarkable work from models and people who get email edits. He groups the fix into altitude, depth and range. Altitude means asking for the outcome rather than the task: not a quarterly deck, but what changed, why, and what to do about it. Depth means working the way the best practitioner would, with the same tools, raw data and methods, and setting targets that feel unreasonable. Range means bringing in other disciplines, including adversarial roles, and asking the model to do the professional's work rather than imitate their style.

The practical core is about verification. He describes making crash reports fetchable, mapping app versions to git commits, and giving Claude access to Windows, macOS and Linux machines so it could reproduce and confirm fixes itself. He also argues for an outside judge rather than letting the model grade its own work. That is the same lesson the transcript data implies: longer autonomous runs are only safe when the agent can check itself against something real.

The caveats Anthropic states

Anthropic lists its own limitations, and they matter. The survey used convenience and purposive sampling, responses were not anonymous, productivity was self-reported and recalled across a year, and the findings describe Anthropic engineers in August 2025, when Claude Sonnet 4 and Claude Opus 4 were the leading models. The study also reports concerns: skill atrophy, a supervision paradox where overseeing the agent requires skills that delegation erodes, and fewer questions flowing between junior and senior colleagues.

The study cites a 67 percent increase in merged pull requests per engineer per day after Claude Code adoption. A merged pull request is a count, not a quality measure, so treat it as consistent with the self-reports rather than proof of them.

For a team outside Anthropic, the honest reading is directional. Expect longer unattended runs if you invest in verification, expect more output rather than shorter weeks, and plan deliberately for the mentoring and review that agents quietly displace. Longer runs also mean the agent rereads the same internal documentation more often, which is where a shared retrieval layer earns its keep.

Frequently asked questions

What did Anthropic reveal about how its engineers use Claude Code?: Anthropic published a July 2025 report of team case studies, a December 2025 study of 132 engineers with about 200,000 Claude Code transcripts, and an October 2026 essay by Felix Rieseberg. They show engineers handing the agent larger tasks with fewer interventions, non-engineers building tools, and verification against real systems as the habit that makes autonomy safe.
How much more productive are Anthropic engineers with Claude Code?: In Anthropic's August 2025 survey, self-reported productivity gains rose from about 20 percent to about 50 percent year over year, and 14 percent of respondents reported gains above 100 percent. Anthropic stresses these are self-reports with sampling and recall bias, and that much of the gain appears as additional output rather than fewer hours.
Do non-engineering teams at Anthropic use Claude Code?: Yes. The July 2025 report describes growth marketing building an agentic workflow that analyzes ad performance and generates variations, the legal team prototyping a phone tree to route staff to lawyers, and product design turning Figma designs into code. Data scientists without TypeScript experience built React visualizations for model performance.
What risks did Anthropic's internal study find?: Engineers worried about skill atrophy as they delegate more, a supervision paradox where reviewing the agent requires the skills that fade with disuse, and less collaboration, since Claude became the first stop for questions once asked of colleagues. Some also described uncertainty about how the profession will change as work shifts toward directing agents.
Can a smaller team copy Anthropic's Claude Code workflow?: The habits transfer, the numbers may not. Start with maintained instruction files, give the agent tests or real systems to verify against, ask for outcomes rather than steps, and review its work with an outside check. Anthropic's results come from engineers with early access to frontier models, so measure your own baseline before and after.

Key takeaways

Anthropic's December 2025 study found maximum consecutive tool calls per transcript rose from 9.8 to 21.2 between February and August 2025, while human turns fell from 6.2 to 4.1.
Respondents said 27 percent of Claude-assisted work would not otherwise have been done, so the gain shows up as more output more than as saved time.
The July 2025 team report shows non-engineering teams, including legal and growth marketing, building their own tools with Claude Code.
Felix Rieseberg's essay argues the bottleneck is ambition and verification: ask for outcomes, give the model real tools, and use an outside judge instead of self-grading.
All three sources come from one company with unusual model access, and the study relies partly on self-reported productivity, which Anthropic itself flags as a limitation.