The OpenAI Codex team breaks it down: from watching ten terminal windows to talking to one long-running manager agent.
The Golden Age of AI Engineering — Alexander Embiricos & Romain Huet & Peter Steinberger, OpenAI · Romain Huet
25 min total·Actually worth watching closely: ~5 min·2 must-watch clips
- 0:13 – 3:45Listen
From code completion to autonomous long-running tasks
Romain Huet opens by tracing how AI coding got here: from completing a few lines, through inline prediction and Command K, to models that take on long, hard goals until they're done. He also notes the change in release cadence — from roughly fifteen months down to about every six weeks.
The real variable isn't any single jump in model capability, it's the frequency of the jumps — at a six-week cadence, the engineering constraints you built last year to work around model flaws are probably liabilities by now.
Mostly the speaker narrating the arc from the stage, with nothing on screen you need to watch — fine to listen to on a commute or while doing something else.▶ Jump to 0:13Speaker · Romain Huet - 3:45 – 7:14Watch
Live demo: the model checks its own work
A live on-stage demo where the model drives the venue's camera and lighting system directly, and shows it can run code and test its own work. The speakers set it against an earlier era when a demo worked maybe three times out of four.
Being able to execute and verify its own results is the dividing line: once that holds, you stop watching line by line, and handing off a whole task becomes possible.
One of the few genuinely live segments — the hardware's actual response is the argument, and most of it doesn't survive in text. Worth stopping to watch in full.▶ Jump to 3:45Speaker · Romain Huet - 7:14 – 12:25Listen
Manage agents like a team, don't use them like a tool
The core mental model: working with agents is just like working with a team — most of the time you talk and let them cook, and only dig into the details on the tricky parts. From there, a critique of existing product shapes: the command line can't carry a collaboration surface for arbitrary work, and the IDE has the order backwards.
The right order is talk about what you want first, then go as deep into the code as you need; opening the file first pins you to implementation details from the start.
Pure argument — the speaker is reasoning rather than demoing an interface, so following the chain matters more than watching the screen.▶ Jump to 7:14Speaker · Romain Huet - 12:25 – 14:08Watch
The open layers of Codex
A walk through the layers Codex was deliberately split into: the harness is open source, forkable, and will run the same agent loop on an open model; OpenAI's own products go through the same open app layer; and what Codex needs gets baked into the same API you use first. The in-app browser and plugins are on screen at the same time.
"The primitives we build for ourselves are the primitives you get" isn't a posture, it's a design constraint — it decides whether you can build your own agent system on the same layer.
The layering and the product surfaces are both up on the big screen; one look saves a paragraph of explanation, and it's one of the densest frames in the talk.▶ Jump to 12:25Speaker · Romain Huet - 14:14 – 18:42Skim
Cost, speed, and where agents run
Alexander Embiricos puts numbers on the efficiency gains: GPT-5.6 Terra brings last-generation-level intelligence at about half the cost, Luna beats some notable models on this eval at $1 per million input tokens and $6 per million output, and Sol runs at around 750 tokens per second on Cerebras. His frame for all of it: value maxing, not token maxing.
Speed isn't about getting the answer a few seconds sooner — it's about letting an agent try five or six approaches in parallel and pick the best, a path that didn't pencil out on cost before and does now.
Mostly pricing tables and eval comparisons — static numbers on screen. Pausing to grab a screenshot of the key figures is faster than listening straight through.▶ Jump to 14:14Speaker · Alexander Embiricos - 18:50 – 22:05Listen
Stop being the human scheduler
Peter Steinberger goes straight at the status quo: juggling ten or more terminal windows, you've degraded into the scheduler, the router, and the memory. The alternative is talking to one long-running manager agent that delegates work to a team of workers. He names the three changes that made it work — server-side compaction making long tasks reliable, coordination letting one thread create and steer multiple projects, and automation triggers.
The bottleneck has moved from tokens, to local compute, to attention — and you can't simply add more attention, so the most valuable skill right now is deciding where to spend it.
The sharpest stretch of the talk, carried by argument and lived experience; there's only the speaker on screen. Good to listen to with your eyes closed, measuring it against how you actually work.▶ Jump to 18:50Speaker · Peter Steinberger - 22:08 – 24:56Listen
Inner and outer loop: you only make decisions
The closing picture of the full loop: an issue gets filed, the manager agent judges it, a worker implements the change and runs the tests, another agent reviews the result, and you review one PR that comes back with the original issue, the proposed diff, and maybe a video. The landing point: coding being abstracted away doesn't mean the end of engineers — engineering is about solving problems.
Models are advancing faster than the harnesses and organizations around them; designing the loop between people and agents is itself the next engineering problem.
The conclusion — the speaker wraps the argument verbally, with no new demos or charts. The last few minutes are enough to get the full position.▶ Jump to 22:08Speaker · Peter Steinberger