An OpenAI engineer's Codex advice: set the checks, decide how much help to add

Eric, who works on the DX team at OpenAI, spent part of a DevDay 2026 session on how to run coding agents. His advice was mostly about decisions, not code. Start small, check the work, add help only when the job calls for it, and keep the goal from slipping. He spoke on the OpenAI show, in the episode "What Comes After Loops?"

What was said

Eric said a long task tempts people to build a whole team of agents. He argued you are almost always better off starting with one, giving it a clear task and a way to check its work. He said the GPT-6 series can keep working, compact its context and push through failures without an outside loop telling it to continue.

Compacting means the agent shrinks its own memory of the job to make room. Extra agents carry a price. In Eric's words: "Every agent Codex brings in can dramatically increase token use without necessarily making the work faster or better." Tokens are the units of text a model reads and writes, and they are what the user pays for.

He described a feature called effort scaling. It lets Codex use stronger or cheaper models for different pieces of a job. He said it can be "a huge help in helping you reduce the decision anxiety in picking the right model for the job."

He also warned about drift. After an agent compacts three or four times, he said, it may move away from the original prompt. His fix was a compact work log holding the overall goal, the running work and how the latest steer relates to both.

Why it matters

Our reading: the scarce skill here is judgement. Someone must define what done looks like, decide whether a second agent earns its cost, and notice when the goal has wandered. Typing the code is not the hard part in Eric's account.

For managers, that suggests budgeting attention for review and for setting checks. Token cost is a decision someone owns, not a background charge. For training, it suggests newer engineers need practice writing tests and goals, not just code.

The other side

This is a vendor presenter at the vendor's own conference, describing his own workflow. Nothing in the excerpts measures how well people judge agent work, or what happens to jobs or training. Those links are ours, not his.

Eric also conceded limits. Effort scaling "won't always cost you less," because even the model that picks cannot perfectly predict the cheapest option. He admitted that some agent setups are not the most efficient way to work token-wise. And he said best practices have a shelf life, so his own advice may age quickly.