Continuous Workspace Agents and GPT-Rosalind Enter Production

OpenAI shifted its enterprise strategy this week with the launch of GPT-Rosalind and Workspace Agents. This dual release pairs a domain-specific reasoning model for biochemistry with a new class of persistent, cloud-based workflow systems designed to execute code and multi-step tasks continuously in the background.

Continuous Cloud Workflows

Workspace Agents operate autonomously in the cloud without waiting for direct user input. Powered by Codex, OpenAI’s code-oriented model, these instances can write code, access internal file systems, and execute scheduled tasks over extended periods. You can configure them to trigger based on specific temporal schedules, like compiling a weekly report, or react to events inside connected applications.

The system launches with integrations for more than 60 enterprise platforms, including Slack, Google Drive, SharePoint, Salesforce, GitHub, Notion, and Atlassian Rovo. These agents utilize shared memory across the workspace context. They update their behavior based on corrections supplied by different team members over time, allowing the underlying model to adapt to internal company conventions and workflows.

Governance and Access Controls

Persistent execution requires strict boundaries. Administrators manage tool access through a new Compliance API, which logs agent history and tracks configuration changes. The system defaults to an always-ask human-in-the-loop requirement for sensitive operations, specifically when an agent attempts to send an email or modify a database record.

Workspace Agents are currently in research preview for the ChatGPT Business, Enterprise, Edu, and Teachers tiers. Usage remains free until May 6, 2026. Following this period, operations will shift to a credit-based pricing model for all executing instances.

Domain-Specific Reasoning

The companion release introduces GPT-Rosalind, a specialized frontier model optimized for genomics, protein engineering, and evidence synthesis. To address biosafety risks, the model is strictly gated behind a Trusted Access program. Usage is restricted to vetted institutions and partners with established internal safety controls, including Amgen, Moderna, Thermo Fisher Scientific, and the Allen Institute.

Early validation indicates a clear advantage over generalized models for biological tasks, tested across established internal and external frameworks.

Evaluation Metric	Comparison Target	Result
LABBench2	GPT-5.4	Outperformed on 6 of 11 tasks
RNA Sequence Analysis	Human Expert Biologists	Scored above 95th percentile

Alongside the core model, a GitHub plugin for Codex connects developer environments to more than 50 scientific databases, including PubMed. This simplifies multi-agent systems that require direct data retrieval from external life science tools for research and experimentation.

Organizational Realignment

The technical releases coincide with a structural shift inside the company. The OpenAI for Science division has been disbanded, and its researchers are now decentralized across other product teams. This reorganization included the departure of three senior leaders: Kevin Weil, Srinivas Narayanan, and Bill Peebles. Concurrently, the company signed a major strategic partnership with Novo Nordisk, indicating a preference for direct enterprise alliances over operating a standalone scientific research division.

If you deploy automated workflows inside a managed enterprise tier, you must review the permissions granted through the Compliance API before enabling scheduled agents. The shift toward continuous execution means your access control strategies need to account for autonomous background actions rather than just real-time user prompts.