Author: abavyka
Coding agent vs chat:
Let’s kick off with the basics: how Claude Code is different from a regular chat with a model, and why some of their problems are the same.
Chat:
You paste a stack trace into the chat, get a fix, paste it into your code, run it, copy the error back again (I think we all started like this) The model can’t see the rest of the project and has no idea whether its code even runs, so some of it is just guesswork.
Agent: model + tools + loop
Tools are actions the model can do on its own: read and edit files, search the repo, run commands, tests, git. The loop means the agent decides the next step itself. The Claude Code docs describe the loop as three phases: gather context, take action, verify the result, and round and round until the task is done.
Examples:
• A test in PaymentService is failing. The agent runs go test ./… or phpunit on its own, reads the code, fixes it and reruns until the test goes green.
•The button doesn’t get disabled during submit. The agent finds the component, makes the change, runs the linter and the tests.
The main thing: the dev stops being a proxy between the model and the code.
Chats have tools too (web search, running code in a sandbox). The difference is where those tools act — in the real repo and terminal — and how many steps the agent takes on its own.
What they have in common
• The same model
• The model remembers nothing. On every step it rereads the whole conversation. Its memory is only what’s sitting in the context right now.
• The context window is limited. It’s all the text the model sees at once. A chat lives within a conversation, an agent within a session.
• Long conversations jumping between topics degrade. The model drops details, mixes up old and new, and fills the gaps with made-up stuff (hallucinations).
• One task — one chat or one session. If the conversation went off the rails, better start over with a short summary.
So why’s it worse in an agent
You write one sentence, and the agent reads five files and runs tests with a page of logs. All of it lands in the context. What takes an hour of conversation in a chat, the agent munches through in 15 minutes of work.
A Claude Code session has /context, so you can check what’s taking up space (system prompt, tools, MCP, CLAUDE.md, messages).
Next up — more on context itself: what goes in there, why a session “goes dumb” and how the cache works.
Read more:
• Coding agent vs chat assistant
• Agentic loop
• Context window
• One task per session
• How Claude Code works (official docs): https://code.claude.com/docs/en/how-claude-code-works
• Anthropic on context engineering: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
• Free Anthropic Academy courses: https://anthropic.skilljar.com. To get started, Claude Code 101 is a good fit, an intro course for folks who haven’t worked with an AI coding agent yet. For those already using it — Claude Code in Action: reading files, running commands, editing code, managing context, GitHub integration.