<![CDATA[No Way Out: Human-Agent Teams]]>https://www.conceptualspiral.com/s/crewshttps://substackcdn.com/image/fetch/$s_!wT3Q!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7185a7fb-6eea-4937-bde5-4d729f0b1709_1024x1024.pngNo Way Out: Human-Agent Teamshttps://www.conceptualspiral.com/s/crewsSubstackMon, 02 Nov 2026 00:57:42 GMT<![CDATA[Turning a Team of Experts into an Expert Team]]>https://www.conceptualspiral.com/p/turning-a-team-of-experts-into-anhttps://www.conceptualspiral.com/p/turning-a-team-of-experts-into-anFri, 30 Oct 2026 20:29:36 GMT

When Eduardo Salas joined me on No Way Out, I asked him to take us back to his work with the U.S. Navy on tactical decision making under stress. He named the question his team was answering at the time: “how do we turn a team of experts into an expert team?”

I’ve had that question in my head all week. On October 29 the DeepMind Institute published “The Economics of Intelligence,” an essay by Alex Imas, Andrew Koh and Shane Legg. The institute says its essays “should not be read as Google’s official view.” Their central claim is that “the individual AI agent is the wrong economic object.”

What the Essay Says

Most benchmarks score one model against one person. The authors argue that “a model can exceed the performance of every person one-on-one and still lose to a well-structured team.”

Their argument has deep roots in economics. Jacob Marschak and Roy Radner, they note, “built an entire theory of teams around the information that is lost in communication alone.” Team science has worked on the same problem from the human side for decades.

The thesis fits in one line: “competence sets the potential productivity level; coordination determines how big an organization can get without becoming too messy to function.” I agree.

How Team Science Came to Me

On July 3, 1988, USS Vincennes shot down Iran Air Flight 655, killing 290 people. In 1990 the Office of Naval Research launched a program called Tactical Decision Making Under Stress, or TADMUS. A paper by Joan Hall Johnston, Janis Cannon-Bowers and Eduardo Salas says the incident “was due, in part, to errors resulting from the ‘Air Warfare’ (AW) team interactions.”

The program worked on more than training. Its objectives covered “decision theory, individual and team training, and information display.” So the Navy went after the decision-support and display side and the team side together.

Salas co-edited the Navy-sponsored TADMUS research volume, Making Decisions Under Stress (Cannon-Bowers and Salas, 1998). On the show he called the program “essentially the Office of Naval Research response” to the Vincennes. His expert-team question is also the title of a 1997 chapter by Salas, Cannon-Bowers and Johnston, “How Can You Turn a Team of Experts into an Expert Team?”, in Naturalistic Decision Making, edited by Zsambok and Klein.

Team science, through Eduardo Salas, emerged from Naval Aviation CRM, built on commercial aviation CRM and lessons from TADMUS.

Long story short, naval aviators who served from the mid-1990s on were trained in DAMCLAS, the seven skills of Navy crew resource management: Decision Making, Assertiveness, Mission Analysis, Communication, Leadership, Adaptability/Flexibility and Situational Awareness. It is arguably the foundation of team science as I learned it. If you are working with agents, call a naval aviator over and have them walk you through the model.

Handoffs and Shared Mental Models

The essay’s picture of a growing team will be familiar to anyone who has coached one. At 20 or 200 people, “implicit assumptions drop out in handoffs between workers.”

Work gets passed along, and context gets lost on the way. In the authors’ words, handing work from one person to another “compresses rich context into lossy summaries.”

Salas told me what gave his team traction with Navy commands: “we use the shared mental model idea.” In The Flow System, which I wrote with John Turner and Nigel Thurlow, a shared mental model is “the development of a collective understanding of, perception of, or knowledge about, a situation or process shared among team members.”

A team with a shared mental model still loses some context at every handoff. It loses less.

The essay asks for “benchmarks for organizations of agents.” One thing they would measure is how well multi-agent systems “share context cleanly across handoffs.” TADMUS studied that problem with Navy air warfare teams.

What I See Missing in Human-AI Agent Teams

Most of the agent work I see goes into making each agent more capable. Very little goes into the roles that hold the team together. The two I see missing most often are boundary spanners and coaching.

In The Flow System, both sit in the distributed leadership helix. Boundary spanners are functional leaders who “operate in the boundaries between teams and between teams and multiteam systems.” They bring resources, foster interactions, coordinate activities and align goals. A coach works toward a team whose members are self-leading and independent.

The essay gets close to this. It says “a skilled coordinator eases handoff friction across colleagues.”

Salas said it his way: “teams live or die by what the team leader does.” He wants team leaders “to be good coaches.”

I told him that when we’re coaching organizations, “we actually want to build internal capacity and capability for that coaching.” That holds for a team of people and agents too. Someone inside the team has to coach the people and the agents as one team.

Orientation Across the Team

The Vincennes keeps coming up in AI conversations. In the October 2026 issue of Proceedings, Tom Hamrick uses it to warn of “AI decisions that are fast, confident, and wrong.”

In the OODA loop John Boyd sketched, orientation shapes what we observe, what we decide and how we act. A team with mismatched orientation can move quickly in the wrong direction. Shared Mental Models are one way a team builds a shared orientation. In the TADMUS work I’ve read, Boyd doesn’t come up, so the connection is mine.

Keith Davids, PhD, who I sat down with recently for No Way Out, comes at teams from ecological dynamics. He and his colleagues describe teams forming synergies through the “direct perception of shared affordances.” That’s a different lens from shared mental models, and I want both on the table for human-AI agent teams.

An organization of agents carries the same risk the Vincennes crew faced. One agent passes along a lossy summary. The next acts on it with confidence. Nobody working the boundary checks whether the team still sees the same situation.

What This Means for Human-AI Agent Teams

Agents are arriving fast. Builders now talk about Boyd’s OODA loop in agent conversations, about loop and graph engineering, and about a chief of staff agent that coordinates the rest. Given that pace, I think it’s likely that human-agent teams can benefit from team science.

We recently used Grok Bots in our Flow Learning Lab, and we can now take human-agent teaming to the next level.

If anything, the need to have agents team is a forcing mechanism, a catalyst to help organizations team.

So here’s a test for this week. Pick one handoff your team made, between two people, two agents, or a person and an agent. Ask the receiver to write three lines: the goal as they understand it, what they plan to do next, and what they think was left out. Have the sender mark every miss, then bring the list to your next debrief. If the misses cluster at one boundary, that’s where to put a boundary spanner or a coach.

References

]]>