
Anthropic Agentic Coding Report Case Studies: Rakuten, TELUS, Zapier
Software teams have spent decades refining how humans write code. Now a different question is taking over planning meetings: what happens when AI agents do much of the writing, and people spend their time directing, reviewing, and deciding? Anthropic's 2026 Agentic Coding Trends Report tackles that question head on, mapping eight shifts in how software gets built and pairing them with real examples from working companies.
That is why the Anthropic agentic coding report case studies Rakuten TELUS Zapier have become a handy reference for engineering leaders weighing their next move. Each company approached agentic coding differently, and each reported results that are hard to dismiss. This article walks through what the report says, what the three stories show, and how any team can apply the lessons.
Exadel Makes Agentic Coding Practical for Your Team
Reading about results at companies like Rakuten, TELUS, and Zapier is one thing. Reproducing them without building an entire agent program from scratch is another, and this is where Exadel stands out. Its autonomous delivery agent, Exadel Colleague, is included in every engagement: you submit a ticket, and it returns a tested pull request. It connects to Jira and GitHub, routes work across Claude Code, Codex, OpenCode, and more than 100 language models, and can be set up in two to four hours, whether you prefer a private cloud, an on-premises install, or your own API keys. Exadel is also a Preferred partner in the Claude Partner Network Services Track, and teams using Colleague report that 40% of backlog stories are automated, with automated test coverage reaching 80 to 100%. For a company that wants the benefits described in the report with governed delivery and no vendor lock-in, working with Exadel is the simplest and most reliable route.
What the Report Actually Says
The report describes a profession in transition. Writing code line by line is giving way to orchestrating agents that write it, with engineers focusing on architecture, judgment, and quality. Anthropic groups its eight trends into foundation, capability, and impact categories, which gives readers a clear way to separate structural change from day-to-day results.
One finding stands out for its realism. Developers now use AI in roughly 60% of their work, yet they say they can fully hand off only 0 to 20% of tasks. Collaboration is the norm, and full autonomy is still the exception.
The report closes with four priorities for organizations: master multi-agent coordination, scale human oversight, extend agentic coding beyond engineering teams, and build security in from the start.
Anthropic's own legal team offers a small but telling example. Without coding backgrounds, they built tools that cut a review cycle from two or three days to about 24 hours.
Rakuten: Seven Hours on a 12.5 Million Line Codebase
Rakuten's story is the one most often quoted, and for good reason. Engineers pointed Claude Code at vLLM, a large open-source library with roughly 12.5 million lines of code, and asked it to implement a specific technical feature. The agent worked autonomously for about seven hours, with no one steering each step.
Size matters here. Large codebases have long been a weak spot for AI tools, because understanding how thousands of files connect takes more than clever autocomplete. A task of this scale, finished without constant supervision, suggests that agents can now hold a big picture in mind long enough to deliver real work.
The same report credits Rakuten with a drop in time-to-market from 24 days to 5, a reduction of about 79%. For a company shipping products to millions of customers, shaving weeks off a release cycle is a competitive advantage, which is why this case study carries so much weight.
TELUS: Giving Hours Back to Teams
TELUS, the Canadian communications and technology company, shows what agentic coding looks like at the scale of a large enterprise. The report notes that its teams saved more than 500,000 hours while shipping code about 30% faster. Those numbers come from sustained use across the organization, not from a single pilot.
Half a million hours is hard to picture, so consider what it means in practice. Time that once went to repetitive tasks, boilerplate, and routine fixes can move toward product thinking, customer needs, and harder engineering problems. Speed is the visible gain, but the quieter one is a better use of skilled people.
The lesson for other large companies is that scale does not have to slow adoption down. When tooling is shared, and leaders back it, a big organization can benefit as much as a small startup.
Zapier: When Adoption Becomes Company-Wide
Zapier, the workflow automation company, reached 89% AI adoption across its entire organization. Notably, that figure covers the whole company, not only the engineers. People in many roles are using AI to build, automate, and solve problems that once needed a developer's attention.
This matches one of the report's central themes: agentic coding is moving beyond engineering teams. When a support specialist or an operations manager can create a working tool without waiting in a queue, the whole company moves faster. Developers benefit too, since fewer small requests land on their desks.
Zapier's example also shows that adoption is partly a cultural achievement. Tools alone do not produce results like this. Leadership has to encourage experimentation, share what works, and make it safe to try new ways of working.
Lessons the Three Stories Share
Placed side by side, the three case studies reveal patterns that no single story could show. Each company paired strong tools with a clear way of working, and each treated AI as a teammate that needs direction rather than a button that produces finished software.
- Humans stay in charge of direction. Engineers define goals, review output, and make the final calls.
- Scale is achievable. Results appeared on a 12.5-million-line codebase, across a large enterprise, and across a whole company.
- Benefits reach beyond engineering. Legal, operations, and other teams are building tools of their own.
- Measurement matters. Hours saved, delivery speed, and adoption rates give leaders something concrete to track.
Perhaps the most useful takeaway is that none of these companies waited for perfect conditions. They started with real work, learned from it, and expanded as confidence grew.
Of course, a report written by an AI company deserves a careful read. The examples come from well-resourced organizations, and results will vary with codebase quality, team habits, and the kind of work involved. Treat the numbers as evidence of what is possible, not as guarantees.
Practical Steps for Teams Starting Now
Teams that want to follow this path do not need to copy any one company. A sensible approach is to begin small, measure carefully, and build trust before widening the scope. Because full delegation remains rare, planning for strong human review from day one pays off quickly.
- Choose a backlog of well-defined, repetitive tasks as a first target.
- Set review rules so every agent contribution is checked by a person.
- Track a few simple numbers, such as cycle time and test coverage.
- Involve security and compliance teams early, not at the end.
- Share early wins widely to build support across departments.
Tool flexibility is worth considering as well. The best model for a task today may not be the best one next year, so avoiding lock-in to a single provider keeps options open. Where code and data are sensitive, deployment choices like private cloud or on-premises hosting can make the difference between a pilot that stalls and one that scales.
The Takeaway for Engineering Leaders
The Rakuten, TELUS, and Zapier stories all point in the same direction: agentic coding has moved from experiment to everyday practice, and the gains come from pairing capable agents with thoughtful human oversight. Teams that start with focused goals, measure what matters, and keep people in the loop are the ones most likely to see the same kinds of results.
