Getting Started

# Getting Started with Agent Coding

## What is agent coding?

Agent coding is a paradigm where an AI agent operates as an autonomous collaborator in your development workflow. Unlike autocomplete (Copilot-style) or chat-based coding (ChatGPT), an agent:

- **Reads** your codebase, docs, configs, and git history

- **Plans** a sequence of steps to accomplish a task

- **Executes** — writes files, runs commands, installs packages, runs tests

- **Iterates** — reads error output, fixes issues, re-runs until the task is done

You describe *what* you want. The agent figures out *how*.

## The mental model shift

| Traditional coding | Agent coding |

| --- | --- |

| You write code line by line | You describe intent, agent writes code |

| You debug by reading stack traces | Agent reads traces and fixes itself |

| You context-switch between files | Agent holds entire codebase in context |

| You run commands manually | Agent runs commands and reacts to output |

| Quality depends on your typing speed | Quality depends on your instructions |

The key insight: **your job shifts from writing code to writing constraints**. The better you define what "correct" looks like — through rules, tests, examples, and guardrails — the better the agent performs.

## Your first session (10 minutes)

### 1. Install Claude Code

```bash

npm install -g @anthropic-ai/claude-code

```

Requires Node.js 18+. Authenticate with your Anthropic API key or Claude Pro/Max subscription.

### 2. Navigate to a project

```bash

cd your-project

claude

```

Claude Code starts in interactive mode, reads your project structure, and is ready to go.

### 3. Give it a real task

Don't start with "write hello world". Start with something you'd actually do today:

```

fix the failing test in pkg/auth/token_test.go

```

```

add a /healthz endpoint that checks database connectivity

```

```

refactor the UserService to use the repository pattern, keep all tests passing

```

### 4. Observe the loop

![Agent Coding Loop](https://claude.ai/chat/img/agent-loop.svg)Watch how the agent:

1. Reads relevant files

2. Proposes a plan

3. Edits code

4. Runs tests or build

5. Fixes any issues

6. Reports what it did

This read → plan → execute → verify loop is the core of agent coding.

## When agents work well

- **Boilerplate-heavy tasks** — CRUD endpoints, migrations, config files

- **Refactoring with clear rules** — rename across codebase, extract interfaces

- **Bug fixes with reproducible errors** — failing tests, stack traces

- **Greenfield features with good specs** — "add X following the pattern in Y"

- **Code review and analysis** — "find all SQL injection risks in this codebase"

## When agents struggle

- **Ambiguous requirements** — "make the UX better" gives bad results

- **Novel algorithms** — agents remix known patterns, they don't invent

- **Large-scale architecture decisions** — they optimize locally, not globally

- **Performance tuning** — they lack runtime profiling context

- **Tasks requiring visual judgment** — UI polish, design decisions

## Key concepts to learn next

| Concept | Why it matters | Page |

| --- | --- | --- |

| [CLAUDE.md](http://CLAUDE.md) | Define project rules the agent follows | [CLAUDE.md](http://CLAUDE.md)[ & Agent Rules](https://claude.ai/chat/claude-md-agent-rules.md) |

| Hooks | Programmatic guardrails that enforce policy | [Hooks & Policy-as-Code](https://claude.ai/chat/hooks-policy-as-code.md) |

| MCP | Connect agents to external tools and APIs | [MCP](https://claude.ai/chat/mcp.md) |

| Headless mode | Run agents in CI/CD without human interaction | [Multi-Agent Patterns](https://claude.ai/chat/multi-agent-patterns.md) |

## Common mistakes beginners make

**1. Being too vague** Bad: "improve this code" Good: "extract the database logic from handlers into a repository layer, add interfaces, update tests"

**2. Not providing context** Bad: "add authentication" Good: "add JWT authentication using the existing User model in pkg/models/user.go, follow the middleware pattern from pkg/middleware/logging.go"

**3. Not verifying output** Agent code compiles ≠ agent code is correct. Always review diffs, run tests, check edge cases.

**4. Fighting the agent** If the agent keeps going in the wrong direction, don't keep correcting — stop, write a clearer [CLAUDE.md](http://CLAUDE.md), add constraints, and restart.

**5. Skipping the [CLAUDE.md](http://CLAUDE.md)** A project without a [CLAUDE.md](http://CLAUDE.md) is like a new hire with no onboarding doc. The agent will guess — and guess wrong.

---

Next: [Tools Landscape →](https://claude.ai/chat/tools-landscape.md)