Working with Coding Agents Begin with the basics

Notes

What changes when the assistant stops suggesting and starts acting.

An agent writes the code quickly. Deciding whether to keep it is still your job.

Begin with the basics Two agents, compared honestly

Where to start

01

The short list of habits

A short list of working habits, each with the failure it prevents.

Read →
02

Two agents, compared honestly

Claude, from Anthropic, and Codex, from OpenAI, are two takes on the same idea: an assistant that lives in your project rather than in a browser tab.

Read →
03

Reading a change properly

The review is the part that was never delegable. It is also the part that gets skipped first.

Read →

Chat has no consequences, an agent does

A chat window is a conversation. An agent in your project opens files, changes them, runs commands and reads the output.

The speed is real, and so is the review it quietly skips.

  1. Mistakes land in the repository, not in a paragraph you can ignore.
  2. The output is a diff, not an answer.
  3. The summary is written by the same thing that wrote the code.
  4. Nothing slows down enough for you to notice you stopped checking.

Before the first session

Most of the ways a first week goes badly are removed before the first prompt.

None of this is about prompting. It is about keeping mistakes cheap.

  1. Work inside version control — a folder with no history is a change you cannot undo.
  2. Start from a clean tree, so the agent's work is its own diff.
  3. Work on a branch you are willing to abandon.
  4. Know the exact command that gets you back to now.
  5. Decide in advance what it may run, not mid-flow.
  6. Keep credentials outside the working tree.

The loop that works

Small steps are not about the agent's ability to hold them. They are about yours to check the result.

  1. State one task with the finish line named.
  2. Ask for the plan before the edits.
  3. Let it work in a bounded step, not on an open mandate.
  4. Read the diff — every changed line, before anything else.
  5. Ask for proof: the output, not “should work now”.
  6. Commit small and often; commits are the checkpoints.

Two failures, then stop

If the same problem survives two attempts, further attempts make the diff worse, not better.

Go back to the last commit and say what actually needs to change. Agents do not get tired, so they will keep layering fixes on a broken approach for as long as you let them.