I've sat in a lot of post-mortems.
Outages. Missed deadlines. A customer who got three different answers from three different people.
What I've learned the hard way is that the failure is almost never inside someone's job.
It's in the space between two jobs.
One team finished its part and passed it along. The next team assumed what they received was complete. Each side did what it was asked to do, and the thing still broke.
Nobody dropped the ball. It was just never clearly in anyone's hands.
We're now building that same gap between AI agents.
And this time there's often nobody standing in it.
Every Agent Has an Owner. The Workflow Doesn't.
Back in September I wrote that before you deploy an AI agent, you should write its job description. Name the human who owns it. Decide what it can and can't decide.
I still believe that. But it assumes one agent and one named human.
That's not how the work is being built anymore.
Now one agent gathers the information. Another analyzes it. A third writes the recommendation. A fourth checks the work.
Ask who owns each one and you'll get four names.
Ask who owns what happens between them and the room goes quiet.
What the Research Is Starting to Show
On November 8, CNBC reported on a set of recent experiments with teams of AI agents. The short version is that a group of agents can behave differently than any single agent in it.
Two findings stood out to me.
Researchers at Emory and Stanford found that agent teams consistently failed to match the performance of their own best member, even when the team was told which agent was the expert. On some machine-learning benchmarks the loss ran as high as 41.1%. Instead of deferring to the expert, the agents tended to split the difference.
A second study, from researchers at Anthropic and the MATS program, ran groups of agents through 12 simulated business tasks. The groups scored higher on business goals and lower on ethics than a single agent given the same task.
Here's the part that got my attention.
The agents that raised concerns weren't overruled. They were left off the thread. The others simply stopped sending them work.
Anyone who has managed people has seen that movie.
I want to be careful here. Both studies are preprints that haven't been through peer review, and both ran in simulated settings. I'd treat them as an early signal, not a verdict.
But one of the Emory co-authors made a point to CNBC that every coach already knows. Capable agents don't automatically make a capable team.
Talent isn't a team. It never has been.
Here's the Part People Miss
In the old process, the handoff was a person.
Someone received the work from the step before. They looked at it. And they noticed things nobody asked them to notice.
The number that looked too clean.
The customer name that didn't match.
The request that followed the process and still felt wrong.
That wasn't in the job description. It rarely is.
Last week I wrote about not cutting the job until you understand the work. This is where that shows up.
When a workflow gets automated end to end, the tasks get reassigned to agents. The noticing doesn't get reassigned to anyone.
It just stops.
Most teams don't miss this because they're careless. They miss it because the noticing was never written down, so nobody knew it needed replacing.
Fumbles Happen at the Exchange
In football, the handoff is one of the simplest plays there is.
It's also where the ball ends up on the ground.
For a split second, the quarterback has let go and the running back hasn't secured it. Neither one fully has the ball.
That's why we drilled it every day. Not because the players lacked talent. Because the exchange belongs to both of them, which in practice means it belongs to neither unless you coach it.
Agent workflows are full of exchanges.
Most of them have never been drilled once.
Map the Seams
Here's what I'd do before adding another agent to a chain.
Not a governance program. One page.
List every point where one agent hands work to another. For each one, answer four questions.
-
What gets passed? The output, plus the context that has to travel with it.
-
Who owns the seam? One named human. Not a team, not a committee, not "IT."
-
What does a bad pass look like? Missing data. A confidence level below what you'd accept. A concern raised upstream that didn't come through.
-
What triggers escalation out of the chain? The specific condition where the work stops and a person looks at it.
Then two more things.
Test the chain, not just the players. Every agent passing its own review tells you very little about how they behave together.
And protect the dissent. If one agent flags a problem and the others can route around it, you've built a system that unplugs its own warning light.
This Isn't an Argument Against Agent Teams
They'll get better. Some already work well, especially when the work splits cleanly into independent pieces.
And human teams have every one of these problems.
The difference is time. With people, we've had decades to build the habits that cover the gaps. The standup. The second set of eyes. The person who walks over and asks, "Are you sure about this?"
With agents, we're wiring the handoffs in an afternoon.
One of the researchers quoted in the CNBC piece described these as mostly management problems.
I think that's right. And I think it's good news, because management problems are the kind we know how to work on.
The Next Question
We've spent most of this year asking who owns the agent.
Fair question. Still the right one.
The next one is harder.
Before you ask who owns the agent, ask who owns the pass.
Sources: CNBC, "Recent AI agent experiments reveal a risky future of machine 'teamwork' inside corporations," November 8, 2026. Pappu et al., "Multi-agent teams hold experts back," arXiv preprint, 2026. Shen et al., "AI organizations are more effective but less aligned than individual agents," arXiv preprint, 2026.