Perplexity launched “Teammates” this week: agents meant to own complex, ongoing work rather than answer one question and forget about it. They keep an eye on a piece of work, follow up as things change, coordinate other agents and hand work to each other. And, the part I find most telling, they can message coworkers in Slack when they need an approval, a clarification or a decision. It’s available to Max subscribers on the web for now, with iOS to follow.

So, the agent does the work, keeps going in the background, and then pings a human when it gets stuck. I can already picture the first message: “Hi! Just circling back on this…” Give it a week and it’ll be suggesting we “take this offline” 🙂

The Bottleneck Moves to the Humans

Jokes aside, I think this launch says something important about where the bottleneck in organisations is heading. When the agent can research, draft, build and monitor around the clock, the slowest part of the system becomes the human who has to say “yes”, “no” or “do it this way instead”. An agent that can finish its part in ten minutes and then waits three days for a sign-off has gained you, well, ten minutes.

I see a version of this already with engineering teams using coding agents. The code shows up in a fraction of the time it used to take, and then it sits in a review queue, or waits for a product decision, or for the weekly architecture meeting. Teams adopt the tools, the activity charts go up and to the right, and the release date doesn’t move. What gives? The work got faster; the decisions didn’t.

Decision Latency Is the New Metric

I believe AI enables velocity, and the companies that still decide by committee (everyone gets a say, the decision waits for the next meeting, the next meeting gets moved) are going to be swallowed by smaller, agentic competitors that decide in an afternoon. Agents like Perplexity’s make that gap visible, because they’ll tell you exactly who they’re waiting on and for how long. That’s a rather uncomfortable dashboard for a lot of leadership teams.

So, if I were running a team with agents doing ongoing work, here’s where I’d focus:

• Decide in advance what an agent can decide on its own. Most “approvals” are about things that are cheap to reverse; write those down as rules and let the agent get on with it.
• Name one person per decision, not a channel or a committee. “The team” never answers a Slack message; a person does.
• Put a clock on it. If a decision the agent needs hasn’t been made within a set time, it escalates, or it goes with a sensible default and tells you.
• Measure how long work waits for a human, alongside how long it takes to do. I suspect most leaders will be surprised by the ratio.

Managing Agents Is Still Managing

Here’s the thing I find quite reassuring in all this: none of the above is new. It’s the same advice I’d give any manager about their people: delegate clearly, give authority along with responsibility, don’t make everything wait for you. Agents simply make the cost of getting it wrong much easier to see, because they never complain, they just sit there, waiting, with a timestamp.

Bottom line: an AI teammate is only as fast as the slowest human it reports to. If you want the velocity, you have to give up some of the sign-offs. Just a thought…