Alexander Embiricos, the OpenAI executive responsible for dots, has made his position clear: he never wants to have a meeting with someone else's AI agent. We think he is right, and the reasoning behind that stance reveals something important about how we should approach agentic AI today. This is not about rejecting automation or clinging to human-only workflows. It is about recognizing that the technology, as it stands, is not ready for that kind of responsibility. When you consider how easily an agent can produce a Create Your First AI Agent in Python With OpenAI's SDK, you also have to consider how easily it can produce a confident wrong answer when no one is watching.
The practical problem is straightforward. If you send an AI agent to a meeting on your behalf, you are trusting that agent to interpret nuance, ask clarifying questions, and know when to stop. That trust is currently misplaced. As we have seen in When Tools Break, Your AI Agent Keeps Writing Confident Answers, agents have a documented tendency to keep generating plausible-sounding outputs even after their underlying tools fail. A broken API call or a misconfigured data source does not trigger a pause; it triggers a hallucination. In a meeting setting, that means your agent could agree to a deadline, commit resources, or misinterpret a stakeholder's concern without any human intervention. The dashboard stays green, but the commitments are wrong.
This is not a knock on the technology's potential. Projects like Explore how four new projects expand agentic AI's reach and security focus show that the ecosystem is actively building guardrails and sandboxed environments. OpenSandbox and Agent Skills are steps toward making agents safer and more transparent. But safety in a controlled environment is different from safety in a live negotiation. Meeting dynamics involve tone, subtext, and the ability to say "I need to check with my team before answering." Those are not skills an agent has yet. The risk is not that the agent will be rude; the risk is that it will be confidently wrong in a way that creates real obligations.
Our takeaway is this: before you let an agent represent you in any interactive setting, you need to be sure it can detect its own failure modes and refuse to act on them. That capability does not exist at scale today. The smart move is to use agents for tasks like data retrieval, note-taking, or summarization after the meeting, not for participation during it. The question to watch is not whether agents will eventually join meetings, but who will build the first one that knows when to stay silent.