Tech • AI • Robotics • Game
What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026
9.4/10TL;DR
OpenAI employees described how long-running Codex agents, paired with Astra and remote compute, generated a detailed Ferrari 3D model, a full Nürburgring fly-through, and autonomously fixed flaky internal CI failures for six weeks.
KEY POINTS
3D modeling from minimal references
A Ferrari exploration project was built from just seven images, producing not only a textured 3D model but also clay and wireframe versions without being explicitly requested. The result was a browsable website designed to inspect the car’s components and construction in detail.
A 21-kilometer Nürburgring render
A separate task generated a complete fly-through of the Nürburgring Nordschleife, the longer 21-kilometer layout rather than the shorter Grand Prix circuit. The render covered all 70 turns and took a little over two days, with most of the time spent waiting on Blender to render frames.
Agents recovered from failures and kept going
During the racetrack build, the system encountered RAM limits, CPU overheating, and other interruptions. It resumed automatically from saved frames and used sub-agents to research the circuit, validate geometry, and plan rendering steps, illustrating how long-horizon agents can continue despite infrastructure problems.
From tightly scripted prompts to goal-based delegation
The workflow has shifted sharply from heavily prescriptive prompting to lightweight outcome-based instructions. Earlier setups required long prompts, code samples, and explicit plans; current models are given a goal, trade-offs, and success criteria, then allowed to determine the implementation steps on their own.
A six-week agent for internal CI maintenance
One internal continuous integration agent ran for six weeks before being stopped, not because it failed, but because other teams had begun deploying similar agents. Its task was to watch for flaky or broken tests, reproduce failures, propose fixes, open pull requests, and coordinate with coworkers.
Output at production scale
Over the run, the CI agent opened more than 100 pull requests per week, pushed over 1,000 commits, sent more than 600 Slack messages, and merged 213 of its own fixes. Across roughly 1,000 hours, it made just under 500,000 tool calls and used more than 5,800 sub-agent sessions, including over 67,000 calls spent waiting on external events.
Sub-agents changed the productivity curve
The CI agent initially worked in a single thread, but after about four days it was restructured so the main thread triaged failures and delegated debugging to sub-agents. Those child workers pulled code, iterated on fixes, reran tests, and pushed changes independently, turning one agent into a coordinated multi-agent workflow.
It solved a nondeterministic GPU bug
One of the hardest cases involved a Heisenbug in floating-point behavior inside a GPU test. The agent correlated logs over weeks, identified that failures clustered on a specific node, taught itself to rerun the test repeatedly there, and then contacted the relevant researchers to determine whether the cause was software optimization or faulty hardware.
Built-in features for long-running work
The setup relied on several Codex features: one thread per outcome, persistent guidance through agents.md, automatic compaction to preserve long context windows, SSH-connected remote environments, Codex Cloud, plugins such as GitHub and Slack, child threads with separate worktrees, Ultra mode for especially hard problems, and scheduled heartbeats for follow-up.
Cost was significant but measurable
The six-week CI run was estimated at about $32 per hour, later dropping to roughly $16 per hour with GPT-6 Sol pricing improvements. The calculation was presented as worthwhile given that the agent repeatedly ran millions of tests and reduced engineering friction across the company.
CONCLUSION
The demonstrations suggest that software agents are moving from short-lived coding assistants to durable operators that can manage multi-day creative tasks and weeks-long engineering maintenance. The central question is no longer whether they can work autonomously for long periods, but which recurring jobs are worth handing over.
Ask a question