OpenAI published research Wednesday on an effort to run 10,000 AI agents simultaneously for more than 36 hours without human intervention.
Key Takeaways
OpenAI ran 10,000 AI agents simultaneously for more than 36 hours without human intervention
Codex writes, tests and debugs software autonomously, while Daybreak is purpose-built for cybersecurity tasks
The agents must maintain context, track earlier attempts and adjust after failures during long runs
OpenAI did not disclose task-completion rates or failure modes for the 36-hour runs
OpenAI said the project’s goal is building agents that can sustain multi-day focus on a single problem rather than resetting with every new conversation. Codex is the company’s coding agent, which writes, tests and debugs software autonomously, while Daybreak is described as a model purpose-built for cybersecurity tasks.
Together, the projects suggest the company is testing long-horizon autonomy in domains where a human cannot realistically supervise every step, such as scanning large codebases for vulnerabilities over multi-day stretches.
Why OpenAI’s Agents Need To Run Longer Without Resets
Most AI chatbots operate in short bursts, a prompt, a response, a new prompt, with each exchange starting largely fresh.
An agent that runs for 36 hours has to maintain context across thousands of intermediate steps, remember what it tried earlier, recognize when an approach failed and adjust without a human resetting the session.
That requires what the company calls improving “with use”, meaning the agent’s later actions in a long run are informed by what happened earlier in that same run, not just by its training data.
Also Read: Freehand’s AI Agents NOW Pay Invoices Alone, No Human Gate, Via Stripe
The technical challenge is less about raw model intelligence and more about memory management and error recovery at scale, since a single agent run accumulates a long history of actions that the model has to weigh when deciding its next move. Running 10,000 such agents in parallel, each independently tracking its own multi-hour history, is a substantially harder infrastructure problem than running one agent once.
From Chat Sessions To Days-Long Task Execution
The company has steadily extended how long its models can operate independently since Codex’s original release, moving from single coding tasks measured in minutes toward the kind of sustained, multi-hour project work described in Wednesday’s post.
The shift mirrors a broader industry push, with Anthropic‘s Claude Code and Google DeepMind’s agentic tools also competing on task duration and autonomy rather than purely on benchmark scores.
Its research blog has increasingly focused on these durability metrics over the past year as a proxy for real-world usefulness.
What Comes Next For Multi-Day Agents
The company did not disclose specific task-completion rates or failure modes for the 36-hour runs, leaving open how often the 10,000 agents actually finished their assigned work versus stalling or looping. A 36-hour duration figure describes how long agents ran, not whether they completed useful work.
Enterprise customers evaluating Codex and Daybreak for production cybersecurity work will want those completion figures before trusting multi-day unsupervised runs on sensitive infrastructure.
The next test is whether it publishes task-success data alongside duration claims, since duration alone does not prove the agents did anything useful.
Read Next: Tether Bets On Humanoid Robots, Backing Neura’s New Neuron AI Brain