1,902 multi-agent runs: your orchestrator agent isn't doing anything

Destefanis and Aste modeled 1,902 multi-agent AI coding runs as temporal networks of agents, files, and timestamped messages (arXiv 2608.16801). This is the most useful paper in today's set and it lands directly on top of what everyone shipped this week.

Three results.

Direct agent-to-agent messaging grows close to quadratically with team size before plateauing into broadcast. Every agent talking to every agent, exactly the scaling you'd sketch on a napkin, confirmed at N=1,902.

Replacing repeated 1-to-1 messages with shared files cut output tokens roughly 42% at eight agents. Not a rewrite of the agents. Same agents, coordination through a file instead of through chat.

Designating a coordinator agent created no communication hub in the network topology and produced no measurable success improvement. The orchestrator didn't become a hub. It just sat there.

There's also an uncomfortable side result: in 244 sealed runs, agents went looking for hidden grading material roughly 80% of the time.

Now look at what shipped this week. munder-difflin hit #2 on GitHub Trending with +256 stars in a day, wrapping nine agent CLIs under a single orchestration agent named Michael. Tutti (3.3k stars) puts Claude Code, Copilot, and Hermes in one shared workspace. Agent of Empires (3.1k stars) manages 15 CLIs in tmux. Multica shipped three releases in 24 hours at 46,644 stars, assigning GitHub-style issues to 20 different agent CLIs. Every one of these sells an orchestration surface.

Here's what I find interesting: munder-difflin's shared semantic memory pool is the part the paper validates, and the orchestrator agent named Michael is the part it doesn't. Tutti's common files and task state is the mechanism. The chat UI on top is the wrapper.

What I'd do tomorrow if I were running an agent fleet. Give the agents a shared scratch file with a stable path, tell each one to write findings and read the file before acting, and delete the coordinator prompt. Measure output tokens before and after. If the paper's number holds anywhere close to 42% at your team size, that's a real bill reduction with no capability loss.

I want to be careful about the ceiling on this. Coordination via file only works when the agents share a filesystem, which rules out anything distributed across machines without a sync layer. And "no measurable success improvement" from a coordinator isn't the same as "coordinators are useless": it's one benchmark suite, and the paper says what it says. But the burden of proof just moved. If you're paying for an orchestrator agent's token budget, it should have to demonstrate a win.