OpenAI published data Friday on how its researchers are actually using coding agents internally. It’s one thing to ship tools for developers. It’s another to show what happens when AI researchers use those same tools to build more AI.

The short version: coding agents are handling more of the grunt work, and researchers are moving faster as a result.

What they’re measuring

OpenAI tracked agent usage across research tasks. The data covers experiment velocity (how fast researchers can iterate), task complexity (what kinds of problems agents can handle), and overall research acceleration.

The post doesn’t give raw numbers, but it’s clear the pattern is holding: agents work well for repetitive, well-defined coding tasks. Things like data preprocessing, running experiments with different hyperparameters, setting up evaluation pipelines. The kind of work that used to burn hours but doesn’t require much creative problem-solving.

What’s interesting is seeing this play out at OpenAI itself. These are the people building the models, and they’re finding enough value to make agent usage a regular part of the workflow.

Who should care

If you’re working on ML research or data science, this matters. OpenAI isn’t the only shop where coding agents could speed things up. The tasks they’re automating (experiment setup, data wrangling, boilerplate) are universal.

For everyone else building software, it’s a decent signal. When the AI lab uses agents for its own research loop, that’s evidence the tools are past the demo stage.

The full post is on OpenAI’s blog. It’s light on specific metrics but useful if you’re trying to figure out where agents fit in a technical workflow.

Also worth watching: Engrim

A new project called Engrim launched this weekend as a universal memory engine for AI command-line tools. It’s built on SQLite and designed to be local-first.

The pitch: most AI CLIs either have no memory or they roll their own storage. Engrim gives you a standardized way to persist context across sessions without sending everything to someone else’s cloud.

It’s early and small, but if you’re building CLI tools that need to remember things (conversation history, user preferences, project context), it’s worth a look. SQLite means it’s portable and doesn’t need a separate database server. Local-first means your data stays on your machine.

The repo has examples and a basic API. If you’re writing AI tooling in Python or Go, the integration looks straightforward.