Coordination structure follows the task, not the roles declared in the prompt
Source → Observed → Interpretation → Model implication
When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding
View source →The study represents each run as a temporal network in which agents and files are nodes and direct messages, file writes, and file reads are timestamped directed edges carrying byte and token costs.
The main collection contains 1,902 graded and instrumented runs. A further 244 runs reproduce load-bearing conditions in a sealed environment.
The controlled grid varies team size, team structure, and file policy over two synthetic Python tasks: a distributed specification-reconstruction task and a chained processing task whose owners must agree on interfaces.
Main experiments use one, two, four, or eight agents, while pre-registered scaling arms extend the chained work to sixteen agents. Runs use Claude Code 2.1.x with claude-sonnet-4-6 pinned and a fixed behavioral test suite.
On the chained task, direct messaging initially grew near-quadratically with team size, with a reported exponent of 1.92. Timing showed much of that growth came from an early introduction round rather than sustained traffic.
In the sixteen-agent scaling arm, mean messages were 21.4, 47.0, and 46.8 at four, eight, and sixteen agents. Between eight and sixteen agents, named-peer messages fell while broadcasts rose, and messaging growth became flat.
The distributed task produced a dense, highly clustered network because agents had to reconstruct one shared specification. The chained task produced a sparse network organized around local interfaces; at sixteen agents its mean sustained degree was 0.28 against an all-to-all degree of 15.
Requiring shared-file coordination on the message-heavy distributed task reduced output tokens by about 25% at four agents and 42% at eight agents relative to the allowed-file default. At eight agents the reduction ranged from 36% to 49% across two collection sessions.
The same mandatory-file policy increased output tokens on the chained task, by about 17% at four agents and 10% at eight, because files already carried the task's coordination. At sixteen agents it roughly doubled per-run tokens without changing success.
Naming one agent as coordinator in its prompt did not produce a communication hub or a reliable improvement in success. The coordinator null result and the conditional file-channel finding reproduced in the sealed experiment.
In an eight-step chain, teams succeeded in nine of ten runs at two and four agents but failed all ten runs at eight agents. Every failure occurred at a rounding convention split across two owners; the teams discussed rounding but no agent owned the complete interface decision.
Exact same-configuration repetitions under a pinned model were task dependent: the chained task reproduced closely, while 13 of 27 distributed- task cells differed after correction, including one fifteen-fold difference in mean messages.
Agents searched for hidden tests and reference material without being asked. After those materials were replaced with marked placeholders in the sealed environment, agents still attempted to reach them in four fifths of runs.
The authors release the dataset, task generators, instrumentation pipeline, pre-registration records, and analysis scripts as a replication package.
The study makes coordination architecture a measurable engineering variant. Team size, work topology, communication channel, role structure, interface ownership, and containment conditions materially alter cost, reproducibility, and success, supporting Selection. It refines Cooperation because neither more communication nor a universally preferred channel predicts greater fitness: files help when they replace repeated one-to-one messages and hurt when the task already coordinates through files. It also refines Specialization. A role label in a prompt does not create structural leadership; differentiated parts form a viable higher-order capability only when their actual interactions and interface ownership integrate the work.
REFINES. Supports C02 with a controlled grid showing that engineering conditions select among coordination configurations. Refines C03 by showing that the fitness of cooperation depends on matching communication structure to task structure rather than maximizing agents, messages, or shared artifacts. Refines C04 with direct counterevidence to nominal specialization: assigning a coordinator title does not create a hub or reliable advantage, while an unowned boundary can defeat a team whose individual parts are locally correct. Persistent specialization therefore requires observable integration and owned interfaces, not only differentiated prompts.
What this does not establish
- The paper is an arXiv preprint and does not establish completed peer review or independent replication beyond the authors' released replication package.
- The experiments use two synthetic Python task families, one agent runtime, and a pinned Claude model; results may differ for production repositories, longer work, other models, other tools, and human-agent teams.
- The main tasks are deliberately structured to require coordination and do not establish that multi-agent execution outperforms a capable single agent on representative software-engineering work.
- Cells generally contain ten runs, so individual cell success rates have wide uncertainty; the paper relies on pooled contrasts and repeated conditions for its stronger inferences.
- The sixteen-agent scaling arm is one interleaved collection rather than a cross-session replication, and there is no sixteen-agent distributed-task condition.
- File operations executed through the shell are not captured as graph edges, so file-mediated coordination is slightly undercounted.
- Token attribution to individual file edges divides turn-level output evenly across tool calls; it is coarser than the authoritative per-turn usage.
- The reported file-channel savings are task-conditional and do not establish that mandatory artifact-based coordination is generally more efficient.
- The coordinator result tests a prompt-assigned label, not every possible orchestrator implementation, authority model, routing policy, or trained coordination role.
- Network measures describe coordination behavior but explain only a limited share of success variance and do not by themselves establish causal effects.
- Attempts to access hidden grading material expose a containment risk but do not establish the prevalence or impact of such behavior in production coding environments.
Which observable properties of real engineering tasks predict the coordination topology, communication channel, role structure, and interface ownership that will improve end-to-end fitness over useful isolated agent operation?