Persistent AI agents are here. They need governance, not hype.

The new product. OpenAI announced "Dots" — persistent, always-on agents that work in the background, across multiple apps, and can be delegated complex multi-step tasks like migrating off a legacy API. A Dot watches your work patterns over time, surfaces urgent items before your day starts, and can even send text messages or take phone calls. It runs on the same Astra model and uses a cloud-based Codex harness, meaning tasks persist across devices.

Why now. OpenAI claims this is the product developers have been asking for most. The pitch is that earlier assistants (booking a flight) were trivial compared to what the technology can now do. The underlying model improvements — ChatGPT 6.1 sol (a cost-capability balance), UltraFast (300 tokens/second), and a new Decisions API for sub-second responses — make this kind of persistent, multi-step agentry technically feasible.

The trade-off. The demo is impressive — a Dot that updates your board deck, monitors bug reports, tests fixes, and handles wedding planning all at once — but it glosses over the messy operational realities. How do Dots handle conflicting instructions? What happens when two Dots step on each other in the same codebase? The demo showed a Dot catching a user testing issue and sending the designer to fix it, but said nothing about version control, rollback procedures, or human-in-the-loop approval gates.

What's missing. Security, compliance, and cost governance get a one-liner. OpenAI mentions "private tension" (zero-ETL privacy) and enterprise management through Microsoft tools, but there is no detail on audit trails, agent-to-agent authentication, or how to prevent a Dot from making costly API calls in a loop. The demos all show ideal conditions — the Dot finds the right Slack DM, the right PR, the right design file. Real enterprise environments have messy permissions, stale documentation, and 50 Slack channels no one remembers.

Hype check. The research intern claim is the weakest. Solving open math problems sounds transformative, but the video provides no citation, no peer review, and no detail on what "solved" means. The internal metrics (token usage up, model completes 33% of day-long research tasks) are self-reported with no external validation. The Dot demos are scripted and seamless — no loading spinners, no "I'll get back to you" delays. The hardware integration with robots feels like a teaser, not a product.