How Anthropic Is Using Claude Code in 2026

Executive Summary

In 2026, Anthropic is not just selling Claude Code — it is running its entire organization on it. Dario Amodei confirmed in late 2025 that on many internal teams, AI is writing “comfortably the majority” of Anthropic’s code, a milestone that has since expanded across the company. Boris Cherny, the creator and head of Claude Code, now writes 100% of his own code via Claude and has not edited a single line by hand since November 2025. This document synthesizes Anthropic’s own published case studies, executive statements, and independent reporting to provide a comprehensive picture of how the company uses Claude Code internally — and what that means for engineers, product managers, and tech leaders everywhere.

What Is Claude Code?

Claude Code is an agentic command-line interface (CLI) that allows developers and non-developers alike to interact with codebases using natural language. Unlike autocomplete-style copilots, Claude Code operates as an autonomous agent: it can navigate entire repositories, write multi-file code, run tests, execute terminal commands, and iterate through feedback loops with minimal human input. It integrates with VS Code, JetBrains IDEs, GitHub Actions, and the web, and supports Model Context Protocol (MCP) servers for connecting to external tools.

Claude Code was originally prototyped as a side project inside Anthropic’s experimental division before becoming one of the company’s most strategically important products. By early 2026, it accounted for roughly 4% of all public GitHub commits, with daily active users doubling month-over-month.

How Anthropic’s Own Teams Use Claude Code

Anthropic published a detailed internal case study based on interviews with power users across ten departments. The pattern is consistent: every team, regardless of technical level, found ways to compress cycles, automate drudgework, and expand what individuals can accomplish alone.

Data Infrastructure

The Data Infrastructure team uses Claude Code to automate routine data engineering tasks, debug complex Kubernetes issues, and create documented workflows that non-technical staff can execute independently. When Kubernetes clusters stopped scheduling new pods, the team fed screenshots of dashboards directly into Claude Code, which walked them through Google Cloud’s UI menu-by-menu, identified pod IP address exhaustion, and provided exact commands to fix it — without involving networking specialists.

Finance team members with zero coding experience now describe their data workflows in plain text — “query this dashboard, get this information, run these queries, produce Excel output” — and Claude Code executes the entire pipeline. The team can now monitor 200 dashboards autonomously, a volume impossible for humans to review manually.

Product Development

The product development team (who build Claude Code itself) uses their own tool to build it — a genuine dogfooding loop. They use “auto-accept mode” (shift+tab) for rapid prototyping, enabling autonomous loops where Claude writes code, runs tests, and iterates continuously. Their most striking example: implementing Vim key bindings for Claude Code. The team handed Claude the entire task as an async job and Claude autonomously delivered roughly 70% of the final production implementation.

For core business logic, the team works synchronously — providing detailed prompts and monitoring in real time for architectural alignment, while Claude handles the repetitive coding work.

Security Engineering

The Security Engineering team uses Claude Code extensively for complex infrastructure debugging. What previously required 10–15 minutes of manual code scanning now resolves in approximately 5 minutes — a 3x improvement — by feeding stack traces and documentation directly into Claude Code. For Terraform infrastructure reviews, engineers paste plans into Claude and ask: “What’s this going to do? Am I going to regret this?” — creating faster security approval cycles that reduce bottlenecks across engineering.

The team also has Claude ingest multiple documentation sources to synthesize runbooks and troubleshooting guides, which become reusable context for debugging future incidents. Notably, the Security Engineering team accounts for 50% of all custom slash command implementations in Anthropic’s entire monorepo.

Inference Team

The Inference team — which manages the memory system underlying Claude’s context window — uses Claude Code heavily to onboard new members and bridge knowledge gaps, particularly for engineers without machine learning backgrounds. What previously required an hour of documentation reading and Google searches now takes 10–20 minutes via Claude, representing an 80% reduction in research time. Engineers also use it for cross-language code translation, generating functionality in unfamiliar languages like Rust without needing to learn them.

Data Science and Machine Learning Visualization

Despite minimal JavaScript and TypeScript experience, the Data Science and ML Engineering teams use Claude Code to build full production React applications for visualizing model performance and training data. In one documented case, Claude wrote a 5,000-line TypeScript application from scratch. The team describes their workflow as treating Claude like a “slot machine” — committing their state, letting Claude run autonomously for 30 minutes, then either accepting the output or resetting fresh.

This approach shifted the team from throwaway Jupyter notebooks to reusable React dashboards for model evaluation — directly improving how Anthropic understands Claude’s own behavior during training.

Growth Marketing

Growth Marketing is one of the most striking non-technical examples. A solo operator uses Claude Code to generate hundreds of ad variations in minutes, compress tasks that previously took 2 hours down to 15 minutes, and deliver what the team describes as “10x creative output”. They use Figma integration and automated ad creation pipelines to operate “like a larger team”.

Legal

Perhaps the most unexpected adoption: Legal team members with no engineering background are now building functional software. Using Claude Code, a lawyer built a working phone tree system and a predictive text accessibility tool in approximately one hour. The team’s internal reaction captures the broader phenomenon: “Holy crap, I’m a developer.”

Key Quotes from Anthropic Leaders

“I think by the end of the year, everyone is going to be a product manager, and everyone codes. The title software engineer is going to start to go away. It’s just going to be replaced by ‘builder,’ and it’s going to be painful for a lot of people.”
— Boris Cherny, Head of Claude Code, Lenny’s Podcast, February 2026

“I think at this point it’s safe to say that coding is largely solved. At least for the kind of programming that I do, it’s just a solved problem because Claude can do it.”
— Boris Cherny, February 2026

“I have not edited a single line by hand since November.”
— Boris Cherny on his own code workflow, Fortune, February 2026

“I think we will be there in three to six months, where AI is writing 90% of the code. And then, in 12 months, we may be in a world where AI is writing essentially all of the code.”
— Dario Amodei, CEO of Anthropic, Council on Foreign Relations, March 2025

“Something that we found is that the value of more senior people with really, really well-calibrated intuitions and taste is going up.”
— Jack Clark, Anthropic Co-founder, on shifting workforce value

“The fun part is figuring out what to build, and coming up with this. It’s talking to users. It’s thinking about these big systems. It’s thinking about the future. It’s collaborating with other people on the team — and that’s what I get to do more of now.”
— Boris Cherny on what Claude Code frees him up to do

Measured Impact: By the Numbers

Team	Before Claude Code	After Claude Code
Security Engineering	10–15 min to debug stack traces	~5 min (3x faster)
Inference	60 min of research	10–20 min (80% reduction)
Growth Marketing	2 hours per ad task	15 minutes (8x faster)
Product Design	Weeks to implement visual changes	Hours to days
Data Science	Disposable notebooks	Persistent 5,000-line TypeScript apps
All Anthropic engineers	Baseline	~200% increase in code output year-over-year

Anthropic confirmed in late 2025 that AI is writing “comfortably the majority” of the company’s code on many teams, with co-founder Jack Clark suggesting the figure could reach 99% by end of 2026. The prediction made by Dario Amodei in March 2025 — that AI would write 90% of code within 3–6 months — appears to have been validated within Anthropic itself, though not uniformly across all teams.

Best Practices from Anthropic’s Internal Teams

1. Invest in CLAUDE.md Documentation

The single most recommended practice across all Anthropic teams is writing detailed CLAUDE.md files — markdown configuration files that Claude reads at startup to understand the project, its workflows, coding style, and constraints. At Anthropic, teams periodically run these files through the prompt improver and add emphasis markers like “IMPORTANT” or “YOU MUST” to increase adherence. End-of-session documentation updates, where Claude summarizes completed work and suggests improvements to the CLAUDE.md, create a continuous improvement loop.

2. Classify Tasks: Async vs. Synchronous

Effective teams develop what Anthropic calls “task classification intuition” — learning to distinguish which work can run autonomously and which needs human supervision:

• Async (auto-accept mode): Peripheral features, prototyping, refactoring, test generation, PR comment fixes
• Synchronous (monitored): Core business logic, critical production fixes, security-sensitive changes, architectural decisions

3. Create Self-Sufficient Verification Loops

The highest-leverage workflow pattern: set Claude up to verify its own work by running builds, tests, and lints automatically after each change. Asking Claude to generate tests before writing code is particularly effective, since Claude can then use the tests to verify and iterate without human intervention.

4. Start from a Clean Git State

All teams emphasize committing to a clean git state before starting an autonomous session and creating regular checkpoint commits. This allows teams to revert if Claude goes off track, making experimental runs low-risk. The Data Science team’s “slot machine” approach — commit, run for 30 minutes, accept or reset — operationalizes this pattern.

5. Use MCP Servers for Sensitive Data

Anthropic’s Data Infrastructure team recommends using MCP servers instead of CLI tools (such as the BigQuery CLI) for data with privacy or security requirements, since MCP provides better logging and access control.

6. Build Multi-Instance Parallel Workflows

Running multiple Claude Code instances in different repositories simultaneously enables true parallel project management. Each instance maintains full context, so engineers can switch between long-running tasks without losing progress or needing to re-explain the situation.

7. Use Screenshots and Multimodal Input

Teams across Anthropic leverage Claude’s vision capabilities by feeding dashboard screenshots, Figma designs, and UI images directly into Claude Code. The Product Design team’s workflow keeps Figma and Claude Code open 80% of the time to enable direct design-to-code pipelines.

8. Prompt for Simplicity

Claude Code’s default tendency is toward complex solutions. Teams explicitly recommend interrupting Claude during a session to ask: “Why are you doing this? Try something simpler.” The model responds well to these corrections.

Risks and Limitations

Security Vulnerabilities

Three critical security vulnerabilities were discovered and disclosed to Anthropic in 2025–2026. The most significant, CVE-2025-59536, exploited Model Context Protocol (MCP) configuration: repository-controlled settings could override user approval safeguards, allowing remote code execution. A second vulnerability, CVE-2026-21852, exposed similar risks. All three have been patched, but the disclosure underscores the structural risk of agentic tools that execute code in developer environments.

Sandboxing Is Off by Default

Anthropic’s own documentation notes that sandboxing — which isolates Claude Code from the broader file system and network — is disabled by default. Many power users intentionally run with --dangerously-skip-permissions to avoid constant approval prompts, which inverts the security model: authority shifts from the user to repository-defined configuration. For organizations handling sensitive data or proprietary code, this requires explicit security policies.

Prompt Injection and MCP Attack Surface

As Claude Code integrates with more external tools via MCP, the attack surface grows. Attackers who can embed malicious instructions in documents, repositories, or data sources that Claude processes can potentially redirect Claude’s actions. Anthropic’s sandbox runtime can sandbox MCP servers, but organizations must enforce this actively.

Limited First-Try Success Rate

Anthropic’s own RL Engineering team reports that autonomous Claude Code tasks succeed on the first attempt approximately one-third of the time. For complex or novel tasks, teams should budget for iteration cycles, checkpointing, and potential full resets.

Tendency Toward Overcomplexity

Claude Code defaults to more complex implementations when simpler ones would suffice. Without active supervision and explicit instructions to simplify, autonomous sessions can produce overengineered solutions that create technical debt.

AI R&D Acceleration Risk

Anthropic’s own Transparency Hub acknowledges that models capable of autonomously conducting AI R&D could accelerate AI progress to a point where risk assessment becomes infeasible. Anthropic tested Claude Opus 4.5 specifically to ensure it does not cross their ASL-4 Autonomy threshold — and confirmed it does not — but this risk category requires ongoing monitoring as capabilities grow.

Implications for Engineers

The most consequential shift is structural, not just productivity-related. Anthropic co-founder Jack Clark acknowledged directly that the value of junior engineering talent inside Anthropic is “a bit more dubious” as basic implementation tasks are automated. Senior engineers, by contrast, are gaining value — specifically those with deep judgment, architectural taste, and calibrated intuition about what to build and how.

Boris Cherny’s prediction — that the title “software engineer” will start to disappear in 2026, replaced by “builder” — is contested, but the underlying dynamic is real: the bottleneck in software development is no longer typing code. It is deciding what to build, reviewing AI output for correctness and security, and maintaining architectural coherence across codebases that are growing at 200% per year.

For engineers navigating this shift:

• Develop taste and judgment, not just technical execution skills. The ability to evaluate what Claude produces — architecturally, securely, and product-wise — becomes the primary skill
• Become a generalist. Everyone on Anthropic’s Claude Code team codes, from the PM to the finance person; the strongest engineers also have product, infrastructure, and business fluency
• Master the review workflow. As code volume triples, the inspection function becomes the critical path. Engineers who can quickly validate, catch edge cases, and reject bad output are more valuable than those who can produce fast
• Invest in CLAUDE.md and tooling infrastructure. The teams with the highest returns are those who invest upfront in documentation, custom slash commands, and agentic loop setups

Implications for Product Managers

Claude Code is dissolving the traditional boundary between product management and engineering. PMs who can articulate problems in plain language and evaluate AI-generated solutions now have a direct path from insight to working prototype — without waiting for engineering resources.

In practice, Anthropic’s experience shows PMs using Claude Code for:

• PRD-to-prototype pipelines: Write a product requirements document in markdown, run Claude Code, and have a functional prototype in 20 minutes
• Codebase exploration: Understand existing systems in plain English using Plan Mode — no code knowledge required
• Data analysis: Process CSVs, generate reports, and synthesize research without writing SQL or Python
• Task orchestration: Replace static to-do lists with context-aware task systems that parse meeting notes, generate Jira tickets, and produce daily briefs
• Writing augmentation: Critique PRDs, generate release notes from commits, and improve stakeholder communications

The implication is that product management is expanding in scope. PMs who can describe what to build, evaluate whether Claude’s output solves the user problem, and catch quality issues will operate at a new level of leverage.

Implications for Tech Leaders

For CTOs, VPs of Engineering, and technology executives, Claude Code in 2026 presents both an opportunity and a governance challenge.

The productivity opportunity is real. Anthropic’s own engineers increased code output by roughly 200% year-over-year as Claude Code became central to their workflow. A team that previously merged 100 pull requests per month may now merge 300 — but the review infrastructure must scale proportionally or defect rates will rise.

The workforce reshaping has already begun. Jack Clark’s framing — that the bottleneck is moving “up the stack” toward judgment and decision-making — has direct hiring implications. Organizations need fewer people executing implementation and more people who can set direction, evaluate output, and architect systems that AI can build reliably.

Security governance requires immediate attention. The combination of agentic autonomy, MCP integrations, and default-off sandboxing means that Claude Code can become a significant security surface without explicit policies. Tech leaders should establish:

• Sandboxing requirements for all production-adjacent Claude Code usage
• MCP allowlists and audit procedures
• Credential scope limitations to minimize exfiltration blast radius
• No-Claude-Code policies for untrusted external repositories

Non-technical team enablement is a strategic lever. Anthropic’s Legal, Marketing, and Finance teams are generating significant output with Claude Code — work that would previously have required engineering resources or been deferred entirely. Organizations that extend this enablement beyond engineering will see disproportionate returns.

The review function becomes the new engineering bottleneck. As Boris Cherny noted: the factory floor is producing 3x as many cars. If the inspection team doesn’t grow — or if it doesn’t develop better inspection tools like Anthropic’s recently released Code Review feature — you ship defective cars at scale. Tech leaders must rethink code review processes, testing coverage standards, and quality gates for AI-generated code.

The Bigger Picture: AI Building AI

The deepest implication of Anthropic’s Claude Code adoption is recursive: AI is now writing the code that trains the next version of AI. As Dario Amodei confirmed, 90%+ of code on many Anthropic teams is AI-written. The company is — in practice — an early prototype of the “AI-native organization” that most tech companies are aspiring to become.

Boris Cherny’s printing press metaphor captures the cultural shift: just as scribes did not disappear when the printing press arrived but shifted from copying to illuminating and binding, engineers are not disappearing — they are shifting from execution to direction, oversight, and taste. The painful part, as Cherny acknowledges directly, is that the transition is not smooth, and the skills that defined a generation of software careers are being repriced rapidly.

Anthropic’s Transparency Hub is explicit that this acceleration carries risks not just for labor markets but for AI safety itself — as AI systems capable of conducting AI R&D could shorten the time available for human evaluation and control. The company’s response is to test each new model against autonomy thresholds and publish risk reports publicly. Whether that governance model scales as capabilities compound is the open question that will define the next phase of the field.

The question worth asking in your next leadership meeting

Anthropic is already a prototype of the organization most tech companies are trying to become. The gap between watching that from the outside and building it deliberately inside your own org is closing faster than most planning cycles assume.

Where in your organization is the bottleneck still execution, and where has it quietly shifted to judgment?