Good coding agent advice

I'm two weeks behind on the newsletter, so I was trying to be responsible by resisting the urge to document the success I've had with my current coding agent setup. My self-restraint has paid off, as Peter Steinberger essentially wrote the exact post I was planning to write.

There's lots of good nuggets in here, and it's uncanny how many I agree with:

1. I also use Codex CLI (well, this fork) on a $200 ChatGPT Pro plan. Claude Code was an epiphany, but their models are overrated for the task, whereas GPT 5's variants are more adherent and diligent across the board. OpenAI's usage limits are virtually infinite by comparison, too
2. I run 3-6 agents in parallel (usually up to 3 per project and up to 2 projects at a time). Unlike Peter, it's rare I let two agents edit the same codebase simultaneously. GPT 5's "medium" reasoning setting is so fast that the time-consuming activities are brainstorming, researching, unearthing technical debt, and planning refactors
3. While git worktrees are a very cool feature, they dramatically slow down code integration with merge conflicts. Additionally, I've found it's hard to avoid API and port conflicts when running numerous development instances simultaneously. And when an environment stops working, agents will silently start coding based on speculation and conjecture. Fast feedback through observable execution of code is the single most important thing, so the risk isn't worth the (marginal) reward
4. Hooks, custom commands, and fancy hacks like coder's undocumented auto-drive mode are nice, but they're no replacement for thinking really hard about what you want

But really, the reason I've had so much success with Codex in comparison to Claude is that if you get off your ass and do the hard thinking necessary to arrive at an extremely crisp and well-informed articulation of what you want, why you want it, and what obstacles it will face, today's agents will generally do a really good job.

Oh, and fuck prompt engineering, just communicate better. As Peter says:

Don't waste your time on stuff like RAG, subagents, Agents 2.0 or other things that are mostly just charade. Just talk to it. Play with it. Develop intuition. The more you work with agents, the better your results will be.

I've started a dozen posts about working with coding agents that I deleted before publishing, because I eventually realized whatever insight I had could just as easily apply to dealing with human colleagues as with LLM agents. Seriously, just talk to it like a human.

Common communication failure modes:

• Telling the agent how to do the work instead of answering why, what, and where, and then getting upset when the ultimate solution manages to complete all the hyper-specific tasks I defined while failing to solve the broader problem
• Giving the agent instructions that contradicted the facts on the ground, only for the agent to spin its wheels endlessly and make a huge mess trying to do the impossible
• Lazily hand-waving away important requirements, only for the agent to miss edge cases it lacked awareness of
• Telling it what I want before I'd really thought things through, then hating whatever it gave me
• Failing to first deal with underlying technical debt, then getting mad when the agent shoehorns in a necessarily-messy solution on top
• Getting frustrated, being condescending, or treating the agent like it's my underling, and then peeking at its reasoning log and seeing 80% of its thoughts are about managing my emotional state and 20% about the problem at hand

I'm more convinced than ever that when people are having a bad time with using AI to write code, it's not only due to ignorance and incompetence surrounding the tools themselves—just as often, it's a failure of imagination and lack of communication skills. Two things that even the best programmers frequently lack. If you're a programmer who's bad at communicating with humans, I hope you've got some other plan for making money in the next decade.

Anyway, that's where things stand in October. June feels like three years ago, so we'll see where we are in February, I guess.