How Coding Agents Actually Work

Everyone thinks coding agents are magic. Type a prompt, get code. Like autocomplete but smarter.

Spend five minutes watching one work and you realize: it's a chaos of subsystems constantly arguing about what to do next.

The Illusion of Intelligence

You ask: "Add authentication with NextAuth."

What you imagine happens: AI understands, writes code.

What actually happens:

1. Router looks at your prompt, decides which model handles it
2. Context retriever scans your codebase, picks "relevant" files
3. LLM generates a plan based on partial context
4. Tool executor runs commands, edits files
5. Orchestrator checks if it worked
6. Loop back to step 3 because it didn't

This isn't intelligence. It's trial and error with better memory.

The Context Problem

Here's what kills most agent tasks:

Agent picks which 4% to look at. Gets it wrong? Writes code that doesn't fit your patterns. Uses libraries you don't have. Ignores conventions.

Why Agents Get Stuck

Watched an agent try to add a feature last week. Loop count: 7.

Each attempt costs tokens. Each costs time. Errors compound because the agent can't see its own pattern of failure.

When Agents Actually Help

Boring, well-defined tasks:

• "Add CRUD endpoints for this model" — works
• "Write tests for this function" — works
• "Rename X to Y across the codebase" — works
• "Refactor this to be cleaner" — disaster

The pattern: agents excel when the solution is obvious and the scope is narrow. They fail when judgment is needed.

The Real Workflow

What works:

Agents are fast interns, not senior engineers. Treat them accordingly.

---

The agent isn't smart. It's just fast at being wrong until it's right.

— blanho