OpenAI Shows How Agents Help Its Researchers and Where They Still Fall Short
In its second report on its internal work, OpenAI says agents handle simple tasks well and hard ones worse. The teams say the key ideas behind the new models came from people.
In its second report, OpenAI describes how they use AI agents to develop their AI models. An AI agent is a program that can complete a task without human supervision. In October, the median researcher at OpenAI used $2,400/day worth of AI agent work at OpenAI's regular prices (a 62% increase over the previous month).
Agents excel at repetitive tasks. They successfully complete 79% of tasks monitoring model training runs. But they struggle with more complex tasks. They only complete 29% of environment construction tasks and 25% of model serving code tasks. Between March and October, the percentage of tasks (2–4 hour duration) completed by an agent without human intervention increased from 23% to 46%. Currently, agents are mostly used for simple, repetitive tasks while humans perform complex engineering tasks.
A survey of 101 OpenAI employees revealed that the median researcher would rather lose 40% of their team's staff than go three months without AI agents. This was an internal, non-representative survey so take this result with a grain of salt. Based on feedback from OpenAI teams, human contributors were responsible for the core concepts behind GPT-6 Astra, a recently released OpenAI model. All data in this post was self-reported by OpenAI and has not yet been independently verified.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles