OpenAI announced in San Francisco on March 20, 2026, that it will transition its development focus toward a fully autonomous AI researcher. This internal shift marks a departure from general-purpose chatbots toward specialized agents capable of solving complex scientific and mathematical problems without human oversight. Company leadership identified this objective as the primary objective for the next several years. Engineers are now tasked with unifying disparate research strands into a single cohesive architecture.
Success hinges on the development of multi-agent systems.
Jakub Pachocki, the chief scientist at OpenAI, recently clarified that the project seeks to move beyond simple text prediction into the area of active discovery. Pachocki previously oversaw the development of GPT-4 and the initial reasoning models that surfaced in late 2024. His current mandate involves creating a system that can formulate its own conjectures, test them through code, and refine results based on internal logic. Unlike previous iterations, these systems will operate over long time horizons rather than responding to immediate prompts.
Pachocki Outlines Timeline for Autonomous AI Researcher
OpenAI established an aggressive schedule for these advancements. By September, the firm intends to debut an autonomous AI research intern designed to handle limited, well-defined problems in math and physics. This intern will serve as a proof of concept for the broader architecture. Internal memos suggest the goal is to test the ability of a model to navigate high-level abstractions without losing logical consistency. The intern acts as the foundation for the final objective.
Full automation is expected by 2028.
OpenAI says that the new goal will be its North Star for the next few years, pulling together multiple research strands, including work on reasoning models, agents, and interpretability.
Documents indicate that the 2028 system will be a multi-agent network. In this configuration, different AI entities collaborate on a single problem, checking each other for errors and improving the final output. Scope of these tasks includes biology, chemistry, and high-level business strategy. Researchers believe this collaborative approach allows for the solution of problems that are currently too complex for human cognition. Each agent in the system focuses on a specific sub-problem, passing findings to a central coordinator.
Convergence of ChatGPT and Codex into Superapp
Separately, internal leaks suggest a radical restructuring of the consumer product line. OpenAI plans to merge its primary tools into a single desktop superapp. This interface will bundle the standard ChatGPT experience with the Codex coding engine and a specialized tool named Atlas. While Codex provides the ability to write and execute software, Atlas functions as an agent-based browser capable of managing the web to gather data and perform actions on behalf of the user.
Integration simplifies the workflow for developers and researchers.
For instance, a user could task the superapp with investigating a market trend and building a predictive model in a single command. The system would use Atlas to find recent data and Codex to build the analysis scripts. Previously, users had to switch between multiple tabs and manually copy data between tools. Consolidation reduces these friction points and creates a unified environment for autonomous agents to operate. Internal testing for the unified interface has already begun in select regions.
Competitive Dynamics Against Google and Anthropic
Pressure from Silicon Valley rivals contributed to this strategic pivot. Anthropic and Google DeepMind have both made significant gains in reasoning and long-context models over the last eighteen months. DeepMind in particular is still a tough threat due to its deep integration with scientific research. OpenAI needs to demonstrate that it can move beyond conversational AI and provide utility in hard sciences. The shift to an Autonomous AI Researcher is a direct attempt to regain the technological lead in the sector.
Capital requirements for this transition are substantial.
Yet, the company maintains that the payoff for a fully autonomous researcher justifies the expenditure. If a model can conduct its own research, the speed of AI development itself could accelerate. It creates a feedback loop where the AI helps design the next generation of its own hardware and software. By contrast, current models are limited by the speed of human researchers who must design every update and evaluate every test result. Removing the human bottleneck is the ultimate goal of the 2028 system.
In fact, early tests show that multi-agent systems can find optimizations in code that human engineers overlooked. These efficiency gains suggest that an Autonomous AI Researcher could far lower the cost of running large models. Still, technical hurdles remain in the field of interpretability. If an AI discovers a new scientific proof, humans must be able to verify and understand the logic used. Without this transparency, the discoveries remain black boxes that scientists cannot fully trust. OpenAI is devoting a major portion of its budget to solving this specific verification problem.
According to Pachocki, the intern model launching in September will focus on this kind of logic-tracing. For one, it will generate step-by-step reasoning logs for every conclusion it reaches. These logs allow human supervisors to audit the internal thought process of the agent. Even so, the sheer volume of data generated by a multi-agent system in the future will likely require another AI just to monitor and verify the work of the primary researchers. The complexity of these systems continues to grow.
The Elite Tribune Perspective
Silicon Valley has long specialized in selling the promise of labor replacement while delivering glorified clerical tools. OpenAI's pivot toward an autonomous researcher suggests a desperate need to find a new narrative for its mounting capital requirements. The current transition from conversational chatbots to agentic researchers is not just a technical evolution but a survival strategy. By framing the future around the 2028 deadline, the company buys itself more time from investors who are growing weary of the high costs associated with simple chat interfaces.
We should view the promise of an AI that does its own peer reviews with deep skepticism. Science is a social process involving consensus, ethics, and physical verification, elements that a text-based multi-agent system cannot replicate. The unified superapp likewise sounds like a move to lock users into a single ecosystem before competitors like Apple or Google can perfect their own OS-level agents. Bundling tools is a classic monopolistic tactic used to mask the slowing rate of improvement in individual components.
If the autonomous researcher fails to deliver a breakthrough in a hard science by 2028, the entire house of cards built on these agentic promises will likely collapse. The firm is effectively betting its remaining credibility on the idea that more compute and more agents can finally replace the human mind.