The Evolution and Future of Agentic Coding with Anthropic's Claude Code: Features, Feedback, and Upskilling for Developers

Anthropic

The Current State of Agentic Coding [0:39]

• Dramatic Shift in the Last Year: Coding has transformed from primarily using an IDE with autocomplete and manual copy-pasting from chat apps to integrating AI agents directly into the development workflow.
  • Previously, AI in coding was limited to basic autocomplete or snippet generation.
  • Now, agents are a fundamental part of the inner development loop, performing text manipulation and even building entire applications with minimal direct human intervention.
• Reasons for the Shift:
  • Improved Models: Models like Sonnet 3.5, 3.7, 4.0, and Opus 4.1 have become significantly better at agentic coding.
  • Enhanced "Harness" (Claude Code): The scaffolding around the model, which includes system prompts, context management, tools, and the interaction layer, has also advanced considerably, akin to a saddle making a difference when riding a horse.
    • Claude Code encapsulates system prompts, context management, tools, MCP server integration, settings, and permissions. This harness provides the necessary structure for the model to perform effectively.

The Evolution of Coding Models [5:20]

• Organic Co-evolution: The improvement of Anthropic's models for agentic coding is an organic process, largely driven by internal "dogfooding."
  • Researchers and engineers building the models use Claude Code daily for their work.
  • This daily use exposes natural limitations and areas for improvement, such as specific types of code edits or the model's ability to operate autonomously for extended periods.
• Learning from Human Intervention: When users, including researchers, have to course-correct or steer the model, these interactions provide valuable data for training the model to perform better autonomously.

Coding Model Evaluation [7:39]

• Challenges with Synthetic Evals: While efforts have been made to build product evaluations and benchmarks like SWE-bench and T-bench, it's difficult to create synthetic evaluations that fully capture the complexity of real-world software engineering tasks.
• Reliance on "Vibe Checks": The primary signal for model improvement often comes from a qualitative assessment by users and developers: "Does it feel smarter?"
  • Personal daily usage across a broad range of tasks (writing new code, fixing bugs, responding to issues, debugging) provides the most comprehensive evaluation.

Claude Code User Feedback Loops [8:56]

• Importance of Responsiveness: A critical factor in building a strong feedback loop was the immediate and active response to user feedback.
  • From the outset, a single Slack channel was established for feedback.
  • The developer (Boris) actively fixed bugs and implemented suggestions quickly, communicating back to users about the changes.
  • This created a positive reinforcement cycle, encouraging continuous feedback and preventing it from feeling like a "black hole."
• User-Driven Development: Given the nascent stage of AI development, listening closely to users is crucial for discovering product direction and identifying necessary features.

The “Hackability” of Claude Code (CLAUDE.md, MCP, Slash Commands) [10:34]

• Design Philosophy: Claude Code was built to be simple and highly "hackable," offering various extension points.
• Extension Points:
  • CLAUDE.md: An early extension point, these markdown files can be placed in root or child directories to provide additional context and instructions to Claude Code, often checked into the codebase.
  • Settings and Permissions System: A sophisticated system for configuring Claude Code's behavior and access.
  • Hooks: An extensive hook system allows users to customize and extend Claude Code's functionality by integrating custom logic at various points in its operation.
  • MCP (My Context Provider): A significant extension point for pulling in external context, such as Slack messages or Sentry logs, enabling more comprehensive debugging and problem-solving.
  • Slash Commands and Subagents:
    • User-defined Slash Commands: These are reusable markdown files within the codebase that define specific workflows, like making commits with predefined instructions and pre-allowed Bash commands.
    • Subagents: Viewed as a type of slash command but with a forked context window, allowing for more isolated and focused tasks.
  • SDK: Anthropic aims to make the SDK more useful for building not just coding agents but any type of agent.
• Future of Hackability: Continued focus on extending Claude Code and making it easier for others to build on top, leveraging ongoing model improvements in autonomy, instruction adherence, and memory.

The Future of Agentic Coding [13:11]

• Hybrid Workflow: The future will involve a mix of hands-on coding (potentially augmented by Claude for text manipulation) and less direct, agent-driven coding.
  • Claude will proactively perform tasks, and engineers will primarily review and approve changes.
• Goal-Oriented AI: Within 12-24 months, Claude is expected to become more goal-oriented, focusing on higher-level objectives rather than specific tasks.
  • Engineers will define monthly goals, and Claude will work towards them, similar to how human engineers break down large goals into smaller changes.
• Higher Abstraction Levels: The evolution will move from Claude making individual file changes to managing entire pull requests and eventually thinking about overarching goals like building complete applications.

How to Upskill for Agentic Coding [14:49]

• Learn the Craft: It remains essential to learn fundamental coding skills, including languages, compilers, runtimes, web app development, program building, and system design.
• Embrace Creativity: With agents handling more of the low-level details, the focus shifts to the idea and the product itself.
  • Coding agents simplify the initial setup and reduce the complexity of modern development stacks (e.g., JavaScript frameworks, build systems).
  • This enables individuals to build out startup ideas or products much faster than before.
• Code is Less Precious: The ability to easily rewrite and iterate on code means the code itself becomes less precious, allowing for more experimentation and flexibility.
• Joy in Coding: While agents automate much of the process, the joy of writing code by hand for fun or specific challenges will still exist.

Claude Code Tips and Tricks [17:49]

• 1. Start with Questions, Not Code [18:03]:
  • For new users, avoid immediately asking Claude Code to write code.
  • Instead, first use it to ask questions about the codebase: "How do I add a new logger?" or "Why is this function designed this way?"
  • Let Claude Code explore the codebase, analyze Git history, and provide answers, building comfort with the agent's research capabilities.
  • Once comfortable with this research mode, then transition to using it for coding.
• 2. Categorize Tasks by Difficulty [18:53]:
  • Easy Tasks: One-shot prompts that Claude can handle almost perfectly.
    • Example: Tagging @Claude on a GitHub issue to generate a pull request directly, freeing up the developer's terminal.
  • Medium Tasks: Start in "plan mode" (Shift + Tab) in the terminal.
    • Align with Claude on a detailed plan first.
    • Once the plan is solid, switch to auto-accept mode for implementation.
  • Hard Tasks: The human developer remains in the driver's seat. Claude acts as a tool for pairing.
    • Use Claude for code-based research, prototyping ideas, exploring system boundaries, or writing unit tests, but the main implementation is done manually.