OpenAI officially launched "dots" on September 29 at DevDay 2026, introducing a new class of "always-on agents" that operate with their own cloud computers and browsers. Unlike traditional LLM interactions that wait for user prompts, these agents perform proactive research in the background with strict read-only tool restrictions. This move aligns with a broader industry trend observed over the past week, including xAI's announcement of Team Bots on September 28 and Meta's Muse for Small Business update on September 29, all converging on event-driven agent architectures.

The Convergence of Event-Driven AI

The technical implication of "dots" is a departure from stateless, request-response models toward persistent, stateful entities. Developer Bobby Hall Jr. illustrates this shift by building a minimal TypeScript implementation that mirrors OpenAI's public description. His model demonstrates how agents can wake up via event feedsβ€”either through polling or push notificationsβ€”rather than waiting for direct user input. The architecture separates background mode, where tools are restricted, from assigned mode, where specific user tasks are executed. This distinction is critical: the event type determines the agent's permissions, not the model's internal reasoning, ensuring that an agent cannot self-authorize high-risk actions during passive research phases.

Authorization as the New Product Surface

The core challenge in deploying always-on agents is not intelligence, but authorization. Hall's implementation highlights a three-tier decision system: actions are either allowed, held for human approval, or denied based on custom rules. Sensitive operations, such as changing passwords, are hardcoded to "always stay with you," mirroring OpenAI's stated policy that certain tasks remain exclusively human-controlled. This suggests that the next frontier in LLM development is not just better reasoning, but robust identity and permission management. As Hall notes, "Always-on is the easy part. Always-allowed is the hard part," emphasizing that audit trails and activity views are becoming essential components of the agent stack.

Key Takeaways

  • OpenAI's "dots" introduce persistent agents with dedicated cloud resources and browsers, moving beyond session-based interactions.
  • A clear architectural pattern is emerging across major AI vendors (OpenAI, xAI, Meta) centered on event-driven wakes and role-specific contexts.
  • Authorization frameworks, including read-only background modes and human-in-the-loop approvals, are becoming the primary product differentiator for agentic AI.
  • Developer implementations suggest that separating event triggers from model reasoning is crucial for preventing unauthorized actions during proactive tasks.

The Bottom Line

The industry is rapidly converging on the understanding that agent autonomy is a security problem, not just a capability problem. OpenAI's "dots" signal that the era of simple chatbots is ending, replaced by complex, permission-bound digital employees that require rigorous governance layers to be viable.

Technical Context

The source material references a specific technical breakdown of this architecture using a mock TypeScript implementation. It highlights the use of an EventFeed class to handle both poll and push deliveries, ensuring the agent reacts uniformly regardless of the trigger method. The implementation also includes a Memory class that segregates user contexts, ensuring that one user's data does not leak into another's session, a critical feature for team-based bot deployments. This practical example serves as a proof-of-concept for the abstract concepts announced by OpenAI, demonstrating that the core logic of "dots" can be modeled with basic event listeners and rule engines.