The future of agentic coding with Claude Code

Key Concepts

• Agentic Coding: Using AI agents to automate and assist in software development tasks.
• Claude Code: Anthropic's tool for agentic coding, leveraging Claude models.
• Harness: The scaffolding around the AI model (Claude), including tools, context management, and system prompts, that enables effective interaction and control.
• CLAUDE.md: A file used to provide additional context to Claude Code, often checked into the code base.
• Extension Points: Features that allow users to customize and extend Claude Code's functionality (e.g., settings, permissions, hooks, MCP, slash commands, subagents).
• Dogfooding: Internal use of a product by its developers to improve its quality and gather feedback.
• Evals (Evaluations): Benchmarks and tests used to assess the performance of AI models and tools.
• Plan Mode: A mode in Claude Code where the user and the AI align on a plan before implementation.
• Auto-Accept Mode: A mode in Claude Code where the AI automatically implements the agreed-upon plan.

The Evolution of AI in Coding

• Past (One Year Ago): Coding involved manual text manipulation in IDEs with basic autocomplete features. AI was limited to copy-pasting code snippets between chat apps and IDEs.
• Present: AI agents are now integral to the coding workflow, handling text manipulation and code generation. Developers interact with agents rather than directly manipulating text.
• Future: The trend of AI agents handling more complex coding tasks will continue, with models becoming more autonomous and capable of achieving higher-level goals.

The Role of Claude Code

• Claude Code acts as a "harness" for Claude models, providing the necessary tools and context management for effective agentic coding.
• The harness includes the system prompt, context management, tools, and integration with MCP servers, settings, and permissions.
• The model and the harness have co-evolved, with improvements in the model leading to better harness design and vice versa.

Co-evolution of Model and Harness

• The development of Claude Code and Claude models is an organic process driven by internal usage and feedback.
• Anthropic employees, including researchers, use Claude Code daily, identifying limitations and areas for improvement.
• This feedback loop informs model training and harness development, leading to continuous improvements in both.
• Examples of improvements include better string replacement capabilities and increased autonomous operation time.

Evaluating Model Performance

• Anthropic primarily relies on "vibes" and real-world usage to evaluate model performance, rather than solely relying on synthetic benchmarks.
• Developers use Claude Code in their daily work and assess whether it "feels smarter" and more capable.
• While product evals are attempted, the complexity of software engineering makes it difficult to create comprehensive synthetic benchmarks.

Internal Feedback Loop

• Claude Code has a strong dogfooding cycle due to a dedicated feedback channel and rapid response to user feedback.
• The creator of Claude Code actively monitors the feedback channel and prioritizes bug fixes and improvements.
• This responsiveness encourages users to provide more feedback, creating a continuous improvement loop.

Claude Code: Current State and Features

• Claude Code is designed to be simple and hackable, with various extension points for customization.
• CLAUDE.md: The original extension point, allowing users to provide additional context to Claude Code.
• Settings and Permissions: Sophisticated systems for configuring Claude Code's behavior and access.
• Hooks: Extensive hook system for integrating custom functionality.
• MCP (Model Control Plane): Integration with MCP servers for managing model behavior.
• Slash Commands: User-defined workflows that can be reused, such as a slash command for making commits with specific instructions.
• Subagents: Similar to slash commands but with a forked context window, providing more isolation and control.

Future of Claude Code

• The focus is on making Claude Code more extensible and easier for others to build upon.
• Improving the SDK to be useful for building coding agents and other types of agents.
• Benefiting from ongoing work to make the model more autonomous, adhere to instructions better, and remember things better.

The Engineer's Role in the Future

• Engineers will likely engage in a mix of hands-on coding and overseeing AI-generated code.
• Hands-on coding may involve guiding AI to manipulate text rather than direct manipulation.
• Claude may proactively make changes, and the engineer's role will be to review and approve them.
• In the future, Claude may focus on higher-level goals, similar to how engineers plan their work over longer periods.

Advice for Engineers

• Learn the fundamentals of coding, including languages, compilers, runtimes, and system design.
• Embrace creativity and use AI tools to quickly prototype and build ideas.
• Recognize that code itself is becoming less precious, and the focus is shifting towards the final product.

Best Practices for Using Claude Code

• Start by asking questions: Use Claude Code to explore the code base and understand its structure and history before writing code.
• Categorize tasks:
  • Easy: Use @Claude on GitHub issues to have Claude write the PR directly.
  • Medium: Start in plan mode to align on a plan with Claude, then switch to auto-accept mode for implementation.
  • Hard: Drive the coding process yourself, using Claude as a tool for research, prototyping, and unit test generation.

Conclusion

The landscape of software engineering is rapidly evolving with the integration of AI agents like Claude. Claude Code serves as a crucial harness, enabling developers to leverage the power of Claude models for various coding tasks. By understanding the fundamentals of coding, embracing creativity, and adopting best practices for using Claude Code, engineers can adapt to this changing landscape and unlock new levels of productivity and innovation. The key takeaway is that the focus is shifting from the process of writing code to the creation of valuable products and solutions, with AI agents playing an increasingly important role in achieving this goal.