OpenAI Researcher Sieb Leaves to Found Company Agent
Maximilian Sieb, a core member of OpenAI's computer operations team, announced his departure from the company to co-found a new company focused on deploying intelligent agents within enterprises. In his post, he mentioned that the team had progressed from making mistakes in booking flights to integrating models into core products, culminating in this week's GPT-6 Astra.
His public resume connects two lines: robotics and consumer intelligent agents, including Covariant for warehouse robotics, Google DeepMind, Character.AI, and then moving into OpenAI's computer usage and reinforcement learning. His personal profile has been updated to "working on new projects," listing his previous employers OpenAI, DeepMind, Character.AI, Mana, and Covariant. The post did not disclose the company's name, funding amount, or product interface, only providing a contact entry for those looking to accelerate real-world deployment.
The timing aligns with Astra's public rollout. OpenAI has positioned computer operations as a core capability of this generation of models: observing screens, clicking controls, and completing multi-step tasks. Sieb's narrative breaks this capability into two parts: solving "can we click correctly" in the lab, and addressing "can it run long-term in company systems" in entrepreneurship. Permissions, data isolation, ticketing systems, and on-call duties are not included in the model demonstrations.
He did not indicate a complete break with OpenAI, nor did he accuse the product of failure. His reasons for leaving are described as a mix of joy and sorrow: bringing computer usage into the mainstream among smart people, then taking the next layer of organizational deployment outside the company. This aligns with a wave of former employees going solo since 2024, with the difference being that he is targeting enterprise workflows rather than retraining another general large model.
In market mechanics, the purchase is based on the expectation that "the model can use computers, but lacks integration into companies"; the sale is the narrative that roles can be replaced solely through ChatGPT plugins. Talent and early funding will flow to teams that can handle SAP, ticketing, and identity directories, rather than just redoing flight booking demonstrations. Beneficiaries are enterprise software vendors needing integrated solutions and former lab founders; those under pressure are agents stuck at the browser plugin level. Funding has not yet seen formal rounds, and pricing power currently lies with individuals, not in ARR.
Computer operations have evolved from booking mistakes to Astra; the next billing will involve permissions and on-call duties, not another benchmarking score.
Source: Public Information
ABAB AI Insight
Covariant sells visual picking to warehouses, Character.AI sells conversations to consumers, and OpenAI sells screen operations into ChatGPT. Sieb is following a similar migration: first getting actions right in a data-closed environment, then moving those actions into paid organizations. The internal milestone for the computer usage team was booking flight dates, indicating that early failures were not in reasoning scripts but in GUI coordinates and state machines. Leaving during Astra's launch week equates to treating "the model can use computers" as a delivered asset while taking the undelivered "which company owns the computer" out the door.
The new company will not mobilize a massive card cluster but will focus on enterprise identities, logs, and accountability chains. The agents will need to modify tickets and handle customer data, with buyers procuring permission matrices and audits, not just another benchmark score. Founders leaving the lab do so because the lab optimizes click success rates, while enterprises need to know who signs off on incidents. Capital will first pay for integration and implementation, then for reasoning.
Benchmarking against Adept's action models as products, Sierra's customer service agents, and a batch of companies embedding browser agents into CRMs. The industry is shifting from "can the model use computers" to "who manages the accounts on the computer." OpenAI continues to sell general computer usage; those leaving sell the integration of general capabilities into specific company directory services.
Structural changes belong to the reconstruction of the industry chain. The mechanism is: once screen operations are usable, the scarcity shifts from pixel recognition to organizational interfaces. Whoever controls employee accounts, approval flows, and data residency can turn agents into roles; those who only control the model can only rent clicks. The next round of company valuations will be based on the depth of residency within customer intranets, rather than the smoothness of flight booking demonstrations.
ABAB News · Cognitive Laws
Once the model learns to use computers, the business becomes about whose computer it is.
Clicking the right button is a research question; entering the organization is a company question.
Labs deliver capabilities; startups deliver accountability chains.