OpenAI Engineers Use AI Agents for On-Call Support

Overview

OpenAI engineers are increasingly relying on artificial intelligence agents to handle high-pressure operational tasks, including after-hours incident response. In a recent demonstration, an engineer named Ali utilized a personalized AI assistant—referred to internally as his "dot"—to manage an on-call emergency at 3:00 a.m.

The incident required immediate triage and investigation. Rather than waking fully to debug complex systems manually, Ali directed his dot to investigate the root cause. The AI agent autonomously launched a Codex thread, allowing Ali to steer the resolution process directly from his smartphone using the ChatGPT application.

This workflow highlights a growing trend in site reliability engineering (SRE): shifting from reactive manual firefighting to supervised autonomous remediation. By leveraging AI agents, engineers can maintain oversight without sacrificing sleep or requiring immediate physical presence at a keyboard.

The 3 a.m. Incident Response

On-call duties are traditionally among the most stressful aspects of engineering work. They require rapid context switching, deep technical knowledge, and the ability to make critical decisions while fatigued. OpenAI’s recent internal use case illustrates how AI agents can mitigate these challenges by handling the initial, time-sensitive phases of incident management.

When the alert triggered at 3:00 a.m., Ali activated his dot. Unlike traditional automation scripts that follow rigid, pre-defined paths, this AI agent possessed the contextual understanding necessary to triage the issue independently. It assessed the severity of the incident and determined the appropriate next steps for investigation.

The dot then initiated a Codex thread. This integration is significant because it bridges the gap between conversational AI interfaces and code-generation environments. Codex, powered by large language models, can understand natural language instructions and translate them into actionable code or diagnostic queries. By starting a thread in Codex, the dot created a workspace where Ali could interact with the system’s infrastructure in real-time.

"His dot investigated the issue and started a Codex thread, so he could steer the work from his phone in the ChatGPT app."

This setup allowed Ali to remain in a state of supervised autonomy. He did not need to manually SSH into servers or navigate multiple dashboards. Instead, he could review the AI’s findings and provide guidance through a familiar chat interface. This reduces cognitive load and accelerates the time-to-resolution, as the engineer can focus on decision-making rather than data gathering.

Steering Work from Mobile Devices

One of the most compelling aspects of this workflow is the ability to manage critical infrastructure from a mobile device. Historically, on-call engineers had to be physically present at a workstation to access internal tools, monitoring systems, and deployment pipelines. This requirement often led to delayed responses when incidents occurred outside of business hours or when engineers were away from their desks.

By integrating AI agents with mobile-accessible platforms like ChatGPT, OpenAI has effectively decoupled incident response from physical location. Ali was able to monitor the progress of the investigation and issue commands via his phone. This portability ensures that senior engineers can provide crucial context and direction without being tethered to a laptop.

The ChatGPT interface serves as a centralized control panel for the AI agent. Ali could ask the dot to dig deeper into specific logs, run diagnostic queries, or draft patches for review. The conversational nature of the interface lowers the barrier to entry for complex troubleshooting, allowing engineers to communicate in natural language rather than navigating obscure command-line tools.

This approach also enhances collaboration. If additional engineers needed to be consulted, the conversation history in Codex and ChatGPT provided a clear audit trail. Team members could join the thread later, read the context, and contribute to the solution without needing a synchronous handoff. This is particularly valuable in distributed teams where time zones may not align with the incident timeline.

Implications for On-Call Culture

The adoption of AI agents for on-call support represents a shift in engineering culture and expectations. Traditionally, being on call has been viewed as a necessary evil, often leading to burnout and degraded quality of life for engineers. By offloading repetitive triage tasks to AI, companies can reduce the mental burden associated with nighttime alerts.

However, this shift also raises questions about accountability and oversight. While AI agents can handle routine investigations, human engineers remain essential for final decision-making, especially in scenarios involving potential data loss or service outages. The model demonstrated by Ali is not full automation, but rather augmented intelligence, where AI handles the legwork and humans provide strategic direction.

Companies looking to implement similar workflows must consider several factors:

1. Security and Access Control: AI agents need access to sensitive systems and data. Robust permission models are required to ensure that agents can only perform actions within predefined safety boundaries.
2. Explainability: Engineers must understand why an AI agent made specific recommendations. Black-box decision-making can erode trust and complicate post-incident reviews.
3. Integration with Existing Tooling: AI agents must seamlessly integrate with current monitoring, logging, and deployment platforms to be effective.

Looking Ahead

As AI capabilities continue to advance, the role of the on-call engineer is likely to evolve further. Future agents may not only triage incidents but also autonomously remediate known issues, requiring human intervention only for novel or complex problems. This could lead to a reduction in page volume and a significant improvement in engineer well-being.

OpenAI’s internal use case serves as a proof of concept for this future. By demonstrating that an engineer can effectively manage an infrastructure crisis from bed, via a mobile app, the company has highlighted the practical benefits of AI-augmented operations.

For the broader industry, this signals a move toward intelligent operations. As organizations adopt similar tools, the standard for on-call responsiveness may raise, but the personal cost to engineers could fall. The key will be maintaining a balance between automation and human oversight, ensuring that AI acts as a reliable teammate rather than a replacement for critical thinking.

The video documentation of this process, referenced in internal communications, provides further insight into the user experience. It shows the seamless interaction between the engineer, the AI agent, and the underlying infrastructure, offering a glimpse into the future of sustainable on-call practices.