OpenAI Speeds Up Incident Management with Ultrafast

OpenAI describes a shift towards "Ultrafast" usage of Codex for managing incidents and engineering tasks. Several members of the technical team report a reduction in response times, to the point of delegating daily operations and considering more ambitious refactorings.

Shortened Response Times Expand Tasks Delegated to Codex

The reduction in wait times with Ultrafast is changing the types of tasks assigned to Codex at OpenAI. Refactorings that were previously postponed are now routine tasks delegated to the AI. Engineers mention working "at the speed of questions," allowing them to follow up on a response, try another approach, and continue without losing track. Adam Perry maintains a continuous Codex thread, which he considers akin to a chief of staff, to which he communicates each morning his appointments and the messages he hasn't yet responded to. He explains that this thread has already analyzed the necessary data to draft a summary, enabling him to ask follow-up questions without waiting for Codex to fetch more information. Sayan Sisodiya now entrusts Codex with manual tasks in the browser, such as filling out forms, which he previously avoided due to delays of about five seconds between each action. He attributes this change to recent improvements in models for computing use and believes the speed is sufficient to let Codex click in his place. Jeff Niu adds that this speed makes large refactorings more accessible, deeming code production very fast and some rewrites almost instantaneous.

Incident Procedures: A Default Mode and Measured Gains

OpenAI engineers indicate that Ultrafast has become the default mode for incident response. AJ Trbojevic explains that security investigations used to take one to two hours to gather, normalize, and contextualize data, whereas they now take 10 to 15 minutes, sometimes almost in real-time. Adam Perry specifies that he no longer systematically duplicates checks in parallel, as Ultrafast's speed allows him to stay focused without having to divide his attention.

30-Second Threshold: Staying in the Workflow

Jeff Niu sets a threshold: if an AI response takes more than 30 seconds, he checks Slack or consults a colleague. He estimates that most of his requests now fall below this threshold with Ultrafast. He describes an interaction that has returned to real-time, preventing him from losing track after a task switch. Adam Perry emphasizes that with equal intelligence, a response in one minute rather than one hour allows for much more to be accomplished.

Tools and Sources: Investigating in Parallel in Codex and Slack

Ultrafast is presented as a high-speed mode in Codex, used to simultaneously examine multiple hypotheses during an incident. Aaron Friel describes opening a dedicated Slack channel and soliciting Codex to monitor, investigate, and respond to questions. He stresses the need to explore many avenues in parallel, which Ultrafast enables. Sayan Sisodiya extends the investigation to Slack to spot references to suspicious changes or symptoms reported elsewhere. Adam Perry and Sayan Sisodiya are part of the core agents team, which develops the components that allow Codex to act, execute commands, and use tools.