tech
How AI coding agents work—and what to remember if you use them
From compression tricks to multi-agent teamwork, here’s what makes them tick.
TL;DR
- AI coding agents use large language models (LLMs) to write, test, and debug software.
- LLMs are pattern-matching machines trained on vast amounts of text and code.
- Agents link LLMs together, with a supervisor assigning tasks to parallel LLMs that use software tools.
- Web-based agents run in sandboxed cloud containers, while local agents require conditional permissions.
- LLMs have a limited 'context' window, affecting their ability to process large amounts of data.
- Techniques like context compression and outsourcing tasks to other software tools help overcome context limits.
- External documentation files (e.g., CLAUDE.md, AGENTS.md) help agents maintain critical context.
- Multi-agent architectures use an orchestrator-worker pattern for complex tasks, but consume more tokens.
- Developers are still responsible for code quality and should use AI agents with good software development practices.
- Human planning and oversight are crucial; 'vibe coding' without understanding is dangerous.
- Studies suggest AI coding tools may not always speed up experienced developers, especially on large, mature codebases.
- Current ideal uses include proof-of-concept demos and internal tools, requiring human accountability.