The ultimate goal for OpenAI is to develop an AGI capable of making novel scientific discoveries, with information synthesis (e.g., writing a literature review) being a key prerequisite.
The breakthrough that enabled truly useful AI agents was the success of reinforcement learning algorithms on complex technical domains like math, physics, and coding.
Significant safety challenges, especially those concerning access to private user data like GitHub repositories, are the most substantial blockers to shipping more capable AI agents.
Users are willing to tolerate longer wait times (e.g., 5-30 minutes) for an AI-powered task if the result is comprehensive and saves them hours or days of equivalent human effort.
Future AI agents must have the ability to access and perform research over private data sources, such as internal documentation and codebases, to be maximally useful.
No Priors Ep. 112 Appearance
Fulford introduced the 'Deep Research' agent, detailing its capabilities (browsing, Python), performance (5-30 mins for hours of work), and lower hallucination rates. She framed the long-term AGI goal as scientific discovery and flagged safety around private data as a major blocker for future agents.
GPT-5 and Agents Breakdown Appearance
Fulford provided broader context on OpenAI's agent strategy, linking the success of reinforcement learning on technical problems to the viability of agents. She also offered insights into OpenAI's internal operations, including its integrated team structure, use of synthetic data for training, and creation of custom evaluation benchmarks.
▶The Agentic Future of AIApr 2026
Fulford outlines a clear roadmap for OpenAI's agents, moving from information synthesis to taking actions and calling APIs. This vision is underpinned by breakthroughs in reinforcement learning and the integration of tools like browsers and terminals, enabling agents to perform most tasks a human can do on a computer.
This signals a strategic shift from conversational AI to autonomous systems, suggesting future value will be in AI's ability to *do* things, not just *say* things, which has major implications for software and service automation industries.
▶Deep Research as a New ParadigmMar–Apr 2026
Fulford details the 'Deep Research' agent, which can perform tasks in minutes that take human experts hours. It achieves this through comprehensive browsing, Python execution, and a lower hallucination rate, positioning it as a tool for serious analytical work.
The tolerance for a 5-30 minute wait time for a high-quality result validates a new product category between instant search and multi-day human analysis, opening up markets in scientific, financial, and business intelligence.
▶Pragmatic Approach to Safety and DeploymentMar–Apr 2026
Fulford emphasizes that significant safety challenges, particularly around private data access, are the primary blockers to releasing more capable agents. OpenAI's approach is described as conservative, requiring user confirmation for irreversible actions and prioritizing safety over rapid feature deployment.
The pace of agentic AI deployment will likely be dictated by solving complex safety and privacy problems, not just by raw capability improvements, potentially leading to longer and more deliberate development cycles than market hype might suggest.
▶Integrated and Iterative Model DevelopmentApr 2026
Fulford describes OpenAI's development process as highly integrated, with research, product, and engineering teams working closely. This structure facilitates rapid iteration, such as using agent-created datasets to improve reasoning models and creating custom internal benchmarks for new capabilities like slide deck creation.
This integrated structure creates a powerful feedback loop between foundational model research and specific product goals, potentially accelerating OpenAI's competitive advantage in building practical, capable AI applications.