The future of agentic coding with Claude Code (AI column)

I saw something interesting. interview Boris Cherny, the creator of Claude Code, was taken from him by Alex Albert. (Claude Relations). They discussed the current state and future of agent programming based on the speed of evolution of models, as well as improvements to the Claude Code product since its inception. As for the key ideas, here they are.

1. Changes in a year are like a revolution. Boris noted the drastic changes in the approach to programming: “A year ago, programming was completely different.” If earlier developers worked with IDE, auto kit and copying code from chat applications, now agents have become an integral part of the workflow. 2. The concept of a harness for an AI model One of Boris’s central metaphors is comparing working with an AI model to riding a horse. “If you ride a horse, you need a saddle. And this saddle is of great importance.” Under the "team" refers to the entire complex of Claude Code: system patches, context management, tools, the ability to connect MCP servers, settings and permissions. 3. Organic Coevolution of Model and Product Boris explained that the improvement of the model comes through dogfooding: At Anthropic everyone uses Claude Code. And that includes researchers.” This allows you to identify natural limitations and improve the capabilities of the system based on real experience. 4. Approach to productivity assessment Instead of formal metrics, the team relies on “whibas,” a subjective sense of improvement. Boris admits: 'Frankly, it's just vibes. Does the model seem smarter? This approach is due to the complexity of creating synthetic tests that would cover the complexity of software engineering, although many would expect a more formalized method of measurement from the creators of the frontier models:) 5. Extensibility as a basic principle Claude Code was originally designed as an extensible system. Among the key features of customization: CLAUDE. md (project context), hook system for integrations, slash commands, subagents with isolated context windows, MCP servers 6. The Future of the Developer Profession According to the forecasts of Boris, 6-12 For months, the developer’s work will include two main areas: more practical programming. (but using Claude to manipulate the text) It also includes less direct programming, where Claude proactively proposes changes and the developer decides to implement them. 7. Learning Philosophy in the AI Era Boris stressed the importance of preserving fundamental skills: "You still need to learn the craft - programming languages, compilers, rantimes, web development, system design." But now creativity and the ability to quickly implement ideas come to the fore. The key change is that the code is no longer "sacred." The ability to re-write with agents shifts the focus from the creation process to the final result. ||Interesting. 1-2 Will we still need to know the basics? Or will agents be so autonomous that most people will forget their knowledge and skills, as most people have forgotten about assemblers?|| 8. Practical advice on the use of Boris recommended a two-step approach for beginners: Don’t write code right away – start with codebase and architecture questions Go to code writing, but consider the complexity Simple tasks: one promptom Medium Tasks: Start with Planning Mode, Agree on Plan, and Then Give Claude Tough Tasks: Stay Lead and Claude Lead as a Research and Prototyping Tool

To sum up, we’re seeing AI tool makers transform the industry from augmentation tools to full-fledged development agents. This requires rethinking workflows and approaches to training new professionals. Well, they run this model on their own development processes, effectively piloting their own product. This is what awaits others in the future:)

#AI #Software #Engineering #Architecture #Agents #ML