---
title: "OpenAI's performance agents show why fixes need a second proof loop"
description: "The Performance Factory tracks over 1,700 investigations, but its strongest design choice is the post-deployment check. Agent workflows should close only when telemetry confirms recovery."
url: "https://devencelab.com/insights/2026/11/20/openais-performance-agents-show-why-fixes-need-a-second"
date: "2026-11-20"
section: "Insights"
tag: "Agentic AI"
author: "Devence Lab"
reading_time: "2 min read"
site: "Devence Lab"
license: "Readable and quotable with attribution to the canonical URL."
---

# OpenAI's performance agents show why fixes need a second proof loop

The Performance Factory tracks over 1,700 investigations, but its strongest design choice is the post-deployment check. Agent workflows should close only when telemetry confirms recovery.

OpenAI says its Regression Workflow has tracked more than 1,700 performance investigations in just over three months and found root causes in over 600. The workflow does not stop at a plausible diagnosis or a proposed code change. After an engineer deploys a fix, an agent checks whether performance returned to baseline.

That final check addresses a common gap in operational automation: completing the requested action is not evidence that the service recovered. The same incident record stays open across investigation, review, deployment and verification, allowing the next agent or engineer to pick up the task with its findings intact.

## Keep evidence with the task

OpenAI describes a task service that tracks state and allowed transitions, a dispatch loop that assigns ready work, cloud execution with access to repositories and observability tools, and subscriptions that wake an agent when a review comment or deployment arrives. This makes work resumable without leaving a model session running while it waits.

The task record matters because agents can reach a convincing but wrong conclusion. OpenAI says an adversarial reviewer checks the diagnosis by testing counterfactuals. In one example, the suspected file-location service had recovered while image downloads still failed; the reviewer reopened the investigation because the theory did not explain the remaining symptoms.

> A fix is complete when the service recovers, not when the pull request merges.

## Make verification a separate stage

OpenAI reports that a manual review of 100 proposed root causes rated 70% accurate, with 30% needing refinement or lacking enough evidence to establish causality. That figure is a useful warning against treating a fluent diagnosis as a reliable incident closure. Separate agents can investigate, challenge the explanation and propose a fix, each leaving evidence for review.

The engineer still decides what ships. OpenAI notes that a proposed cache might improve latency while introducing stale data, a trade-off that needs service context. The workflow therefore automates evidence gathering and proposal while preserving human review at the decision with production impact.

## Close the loop with a measurable baseline

For teams building similar systems, define the recovery signal before enabling the agent: affected request groups, latency or availability measures, and the time window that counts as restored. After deployment, compare those same signals with the pre-incident baseline. If only part of the service improves, record the remaining regression as a new task; if nothing improves, reopen investigation.

OpenAI also says it runs a daily agent to find recurring tool failures and access errors, then creates repair tasks for repeated problems. That extends the loop from fixing product regressions to fixing the automation itself. The operational lesson is straightforward: make every workflow wait for external evidence before it declares success.

## Sources

- [How OpenAI’s Performance Factory keeps ChatGPT fast](https://openai.com/index/how-openai-performance-factory-keeps-chatgpt-fast/) - OpenAI
