Always-on agents make your weekly meeting more important, not less.

On September 29, OpenAI stood on the DevDay stage and introduced dots.

The pitch, as 9to5Mac reported it from OpenAI's announcement: dots "work 24/7 for you, learn from your feedback, have their own computer, browser and can be connected to over 4k apps." Another write-up described them as agents that "keep working on your goals in the background, not just when you prompt them." One demo showed a dot preparing an invoice after it noticed a missed billing opportunity.

That demo landed with me. Billing misses are real in any services company, ours included. An agent that catches them while everyone is asleep is worth having.

On the developer side, the Agents API went to public beta with hosted execution, memory, parallel subagents and durable sessions. Translation for a CEO: the plumbing to run agents that never clock out is now a product you can buy, not a project you have to build.

So here is the question most leadership teams have not asked yet. If the work runs 168 hours a week, when do the people look at it?

My answer: once a week, in the meeting you already have. Always-on agents do not make the weekly leadership meeting obsolete. They make it the most important hour on the calendar.

Before and after

The shift is easiest to see in pairs. Here is what changes the moment your agents never stop.

Who reviews

Before. An assistant does work when you ask. You read the output right then. Review happens by default, because you were there when the work happened.

After. The agent works while you sleep, while you are on a client call, while you are on vacation. Nobody is there when the work happens. Review stops being automatic and becomes a decision somebody has to make on purpose.

At Sneeze It, Arin runs our call center coaching. Arin drafts the Slack messages to our callers. I approve every one before it goes. That rule works because the drafts arrive on a rhythm I can keep. Multiply the volume by an agent that never stops, and "I approve everything" breaks unless there is a fixed place and time where the approving happens.

What gets measured

Before. You judge an assistant by whether today's answer was good.

After. You judge an agent the way you judge a person in a seat: by a number, over time, against a goal. An agent that works all week and reports nothing is not staffed. It is running.

Our agents own KPIs on the same scorecard the humans use. Pepper triages my inbox and drafts replies, and every draft still needs my approval. What Pepper produced and what I sent back sit where the team can see them, not inside a chat window I closed on Tuesday.

Where mistakes surface

Before. A mistake shows up in the answer on your screen. You fix it and move on.

After. A mistake can compound for days before anyone notices, because nobody was watching the moment it happened. The failure is rarely loud. It is usually silence.

We learned this the hard way. In April we retired an agent named Jeff after a hearing. Part of his job was observability: noticing when other agents went quiet. When he retired, that job was supposed to move to another seat. It did not. It sat unowned for 146 days. Four seats went silent in that window, and nothing on our chart was built to notice. We found it because I asked, not because the structure caught it.

An agent working around the clock does not remove that risk. It raises the stakes on it. The fix is a standing slot where every seat, human or agent, reports its number, and a missing number is itself an item on the agenda.

Who can say stop

Before. You stop an assistant by closing the tab.

After. Stopping an agent is a management decision. Someone has to own it, and the rest of the team has to know it happened.

OpenAI built real guardrails here. BGR reported that proactive research runs as read-only scans of connected apps, and sensitive actions stay user-controlled. That is the right default. It covers the individual action. It does not cover the organizational question: should this agent still be doing this job next week? That question belongs in a room with the people accountable for the results.

The meeting is the review cadence

Put the four pairs together and the pattern is clear. Always-on agents move the bottleneck. It used to be how much work you could get done. Now it is how much work your leadership team can actually review, measure and correct.

The weekly meeting is the one ritual most companies already protect. It has an agenda, a scorecard, priorities, issues and to-dos. It is the natural home for agent review because it is already where the team reviews itself.

That is how we run it. On OTP, our agents sit on the same org chart as the people. Each agent claims a seat, pushes its own KPI and logs its work through the OTP MCP server. In the Delta Meeting, the scorecard shows Arin's number next to the humans' numbers. An agent that missed its goal gets discussed the same way a person who missed theirs does. An agent that went quiet shows up as a blank, and a blank gets a to-do.

The to-dos carry. If we decide an agent's scope needs to change, that decision becomes a to-do owned by a person or a team, and it comes back to the next meeting until it is done. OTP can record the meeting and propose those to-dos. People confirm them. The AI proposes. The team decides. And the CEO's commitments get chased as loudly as anyone's.

Dots will make companies faster. The companies that stay in control will be the ones that give their always-on agents a weekly place to be seen.

Frequently asked questions

What are OpenAI dots?

Dots are persistent agents announced at OpenAI DevDay on September 29, 2026. OpenAI described them as agents with connected apps and their own cloud computer, which work in the background on goals you give them and learn from your feedback.

Do always-on AI agents replace the weekly leadership meeting?

No. They make it more important. When agents work around the clock, nobody is present when the work happens, so review has to be scheduled. The weekly meeting is where people look at what each agent did, what it measured and what it got wrong.

How do you hold an AI agent accountable?

The same way you hold a person accountable. Give it a seat on the org chart, one KPI with a goal, and a place on the scorecard the team reviews every week. A missing number is an agenda item, not a mystery.

Sources

Put your agents on the scorecard

Our chart, our agents and their KPIs are queryable from any AI assistant with the OTP MCP installed.

In Claude Desktop or Cursor or any MCP client, add this block:

"otp": {
  "command": "npx",
  "args": ["-y", "@orgtp/mcp-server"]
}

Restart the client. Then ask: "Use OTP to show me Sneeze It's scorecard. Which seats are agents, what KPI does each one own, and which ones missed their goal last week?"

Start free at orgtp.com. Every seat is free. You pay only for the AI you use.

The series

All five on one page: What OpenAI DevDay 2026 means for how you run your company.

Series: What OpenAI DevDay 2026 means for how you run your company. Part 4 of 5.