Using AI / 25 Jun 2026

The coding agent will change. The workflow is what you keep

Most AI tools are improving. Coding agents are improving faster than almost any of them.

The stack that felt settled three months ago is not settled now. Claude Code, Codex, Gemini CLI, open-source agents, remote setups that run a task while you do something else. Each one becomes the right choice for a particular job, and the right choice keeps moving.

It is tempting to pick a side and learn it deeply. That is the wrong bet.

The question is not which agent wins

The mistake is to ask which coding agent will win. The better question is whether your development workflow can survive the answer changing.

A coding agent is fast at producing code. It is not the part that makes the code safe to ship. That part is the system around it: a clean repository, tests that catch mistakes, source control that lets you undo them, and a clear description of what you actually want built.

Swap the agent and that system still holds. Lose that system and a better agent just helps you make a mess faster.

Faster code, faster mess

A stronger agent raises how much one person can attempt. It does not raise how much the surrounding system can safely absorb.

If the tests are thin, more code means more untested code. If the repository is messy, more changes means more confusion. If the task was vague, more output means more of the wrong thing.

The agent got faster. The constraint moved to the engineering around it. For coding the fix is unglamorous and well understood:

Keep the repository clean and its context legible to a model.
Keep tests strong enough to catch what a fast agent gets wrong.
Keep tasks small, so a change is easy to review and easy to revert.
Keep source control tight, so any agent's work is recoverable.

None of that is about a particular tool. All of it makes any tool safer to hand work to.

What stays human

The agent writes the code. Someone still decides what is worth building, judges whether the result is right, and owns it when it ships. That part no tool upgrade replaces.

Prompt you can use

I want to check how well my development workflow would survive switching coding agents.

Here is how I currently work: [describe your repo, your tests, how you give the agent context, and how you review and merge its changes].

Ask me questions until you understand it, then tell me:
1. Which parts depend on the specific coding agent I use today.
2. Which parts would still hold if I switched to a different agent next week.
3. The two or three changes that would make my workflow most tool-independent.

Be specific and practical. Assume the agent keeps getting better and that I will want to switch.

The rule

Do not optimise for the coding agent. Optimise for the workflow it plugs into.

The agents will keep changing, probably faster than anything else in AI. A workflow built on clean context, real tests, small tasks and tight source control does not care which one you use this month. That is the thing worth getting good at.