← All writing

AI Agents

Agent Coordination: Why AI Agent Teams Talk Too Much

Mani G 3 min read

Eight blue AI agent nodes in a ring connected by tangled message lines around a shared message board, next to the words Agent Coordination

Agent coordination is the skill everyone assumes AI agents already have. Spin up eight agents, give them a shared chat, and they will split the work like a good team. A new benchmark says that assumption is mostly wrong. On November 18, Bespoke Labs released the Multi Agent Coordination Bench (MAC Bench), and its main finding is blunt: current frontier agents do not coordinate well.

What agent coordination actually means

Coordination is the part of teamwork that is not the work itself. Who takes which piece, who flags a dead end, and who notices two people just fixed the same bug.

MAC Bench tests this with twelve long tasks built so that one agent's work depends on its teammates. The list reads like a sci-fi syllabus: air traffic control at Dubai, a Mars sample-return mission, power grid restoration, a zombie survival game, and even turning a mass spectrum into a molecule. The tasks are in an open GitHub repo.

Each task runs three ways. One agent alone. Eight agents with a shared message board. Eight agents with no way to coordinate. Every run gets four hours of wall-clock time, and nobody gets a role. Splitting the work is part of the test.

That third setting is the clever bit. It separates the gain from having more hands from the gain from talking.

The numbers

Averaged across Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon:

  • One agent scored 0.35.
  • Eight agents with a message board scored 0.42.
  • Eight agents with no coordination scored 0.44.

So the team beat the solo agent, but only because it had more workers. The talking added nothing on average. Bespoke puts the effect of coordination at minus 0.018, with a 95% interval that crosses zero, so the fair reading is "no proven benefit", not "chat makes agents dumber". As Magica points out, these are Bespoke's own results and nobody has replicated them yet.

A few details stood out to me:

  • Opus 5.5 was the only model whose teams gained from messaging. Alone it scored about the same as GPT-6.1 Sol. In a team of eight it hit 0.53 against 0.41.
  • In teams of just two, 22% of tool calls already went to the message board.
  • When an agent posted a finding a teammate could act on, a teammate changed course within five minutes only 34% of the time.
  • A team of eight cost 8.8 times as much as one agent, and reached the solo score a median of 2.2 times sooner. Not eight times.

A quick example

Picture a team of eight agents fixing a slow checkout service. Agent three finds that the slowdown comes from one database query and posts it. Agents five and seven are deep in their own profiling and never read the board. Twenty minutes later, agent seven proudly reports the same query. Meanwhile agent one reverts a commit based on a stale message. Nobody did bad work. They just worked blind.

That is the failure the benchmark keeps catching. The report logs episodes like "message not acted on", "stale information" and "result not shared". It reads like a retro from a human team, which is either comforting or worrying.

My take

I think this is the most useful agent benchmark of the month, because it measures what vendors hand-wave. "Agent swarm" sounds great in a pitch deck. But if eight silent agents beat eight chatty ones, the smart move today is often boring: give each agent a separate slice of the problem, keep the chatter low, and merge the results yourself. A single agent gave the best score per dollar in every task and model pair.

The surprising part is how old this lesson is. In 1975, Fred Brooks wrote in The Mythical Man-Month that adding people to a late software project makes it later, because communication overhead grows faster than headcount. Fifty years on, we built new workers and handed them the same problem.

So here is my question: if you run multi-agent setups, do your agents talk to each other, or do you keep them quiet and merge their work yourself?