Coordinating embodied agents becomes increasingly difficult as teams grow in size, diversity, and mission duration. Each agent contributes a distinct set of capabilities, executions, and observations, but also introduces additional decisions about task allocation, communication, timing, and resource use. In a small homogeneous team, a single manager may directly assign work to every agent. However, in a large heterogeneous team, the manager must simultaneously track the state of many workers, reason over their capabilities, preserve temporal dependencies, and revise assignments as the environment or information changes.
The Challenge of Organizing Embodied Artificial Collectives
Additionally, these demands are not captured by team size alone. The structure of the work is equally important. Some tasks are naturally divisible. Multiple drones can search different regions or multiple firefighters can remove different obstacles concurrently. Other tasks contain prerequisite relationships. A fire must be located before crews can be transported, helicopters must finish their previous transportation task to empty the load before taking more assignments, and access routes need to be identified before ground units can be deployed safely.
Complex missions often combine both forms. They can also involve dynamic changes such as unexpected environmental changes, agent or communication loss, or new information becoming gradually available. An artificial organization must therefore coordinate parallel work within stages while preserving dependencies between stages and adapting the overall coordination.
The challenge is therefore not only to produce a better plan. It is to determine an organizational structure that matches the mission, the workforce, and the relationships among the required activities. This requires reasoning about which agents should be grouped, which managers should supervise them, how many levels of management are warranted, and what form of coordination should govern each group.
A Testbed for Organizational Intelligence
We studied artificial organization in an extended version of CREW-Wildfire, an embodied environment originally developed to evaluate large-scale embodied collaboration among language-model-based agents. The original environment introduced procedurally generated wildfire environments together with 12 task levels, including tree removal, scouting, transportation, civilian rescue, fire suppression, and integrated wildfire response, with small and large variants yielding 16 benchmark configurations.
We expanded this benchmark to 25 missions designed to span a broader range of organizational demands, including larger teams, heterogeneous workforce compositions, longer planning horizons, multi-stage objectives, and unexpected changes during execution.
CREW-Wildfire operates in procedurally generated landscapes with seed controls for reproducibility. Continuous fields describing elevation, moisture, and wind are generated together with discrete terrain and land-cover types such as forest, brush, land, and water. Settlements and other mission-relevant features are placed according to the scenario seed. These variables alter the spatial distribution of hazards and resources and affect the evolution of the fire, such that the same high-level objective can require different search, transportation, and intervention strategies.
The wildfire evolves dynamically through a cellular-automata process whose local propagation depends on environmental conditions including terrain slope, vegetation, moisture, and wind. Fire can therefore expand in different directions and at different rates as a mission unfolds.
For the organizational problem studied in this work, this dynamic is consequential. Observations collected early in an episode can become outdated, access routes can become unsafe, and the relative urgency of reconnaissance, transportation, rescue, containment, and direct suppression can change over time. Teams consequently need to coordinate not only across space but also across an evolving physical process.
CREW-Wildfire contains four worker types with complementary capabilities:
- Firefighters are general-purpose ground agents that can remove vegetation, rescue civilians, refill water, and directly suppress fire.
- Bulldozers clear vegetation efficiently but cannot rescue civilians or extinguish fires.
- Drones provide rapid wide-area reconnaissance but lack physical intervention capabilities.
- Helicopters provide long-range mobility, can transport firefighters, and support water-based operations.
These extensions were designed to make the benchmark particularly suitable for studying organization. For example, a sparse tree-removal mission primarily tests whether work can be partitioned among independent agents. A helicopter-firefighter transportation mission requires complementary roles and synchronized handoffs. Search-and-rescue missions make later actions depend on information obtained during reconnaissance. Fire containment combines ground workers with different intervention capabilities.
The most complex suppression missions require reconnaissance, resource acquisition, rescue, transportation, deployment, firebreak construction, and direct suppression to be coordinated over time. Thus, the 25 tasks vary not only in conventional difficulty but also in the structure of coordination they demand.
Task-Specific Organizational Hierarchies
Given a task description, the mission requirements provided to the organizational system, a set of available workers, the embodied agents available for assignment, and their capabilities, the actions and roles each worker can perform, ORCH constructs a rooted organizational hierarchy represented by a tree.
Its leaf nodes correspond to the available embodied workers, while other nodes correspond to managers. The hierarchy can be designed by a human expert or generated by a language model.
Each hierarchical team organization is task-specific. Simple missions with a small number of workers can retain a shallow structure in which one manager supervises all workers. Larger missions can produce a team of teams in which a root manager supervises several intermediate managers, each responsible for a coherent group of sub-teams or sub-goals.
The generation process is encouraged to keep the organization as simple as possible while introducing additional decomposition when one manager would otherwise need to coordinate too many independent workers or when the mission contains distinct interdependent sub-goals.
Once constructed, the organizational structure remains a scaffold for execution. One can potentially explore dynamically changing this structure during task progression. This study takes the initial step toward incorporating human organization theory, so such exploration is left as future work.
Though the structure holds for the given task upon construction, the active plan, phase, and assignments can change as the mission progresses. Thus, adaptation during execution can still occur through the information, decisions, and consequences that move through the organization.
Closed-Loop Execution Through Hierarchical Communication
To keep a team of teams aligned as the environment and mission state evolve, ORCH continuously exchanges local progress upward and updated decisions downward through the hierarchy. It alternates between a bottom-up status phase and a top-down planning phase at each environment step.
During the top-down phase, decisions propagate from the root toward the leaves. Horizontal managers directly distribute concurrent assignments among their children. Vertical managers first evaluate the active phase and the sequence of future phases, and then assign work needed for the current phase.
Before the plan is finalized, children evaluate whether their assigned roles are feasible and consistent with their capabilities. Actionable objections can be returned to the manager, who revises the plan until no further conflict is identified. Workers then translate their accepted assignments and current observations into environment actions, which are executed simultaneously.
This bi-directional protocol provides each manager with a compressed view of the state relevant to its responsibilities. Workers do not need to reason over the full communication history of the collective, and the root manager does not need to directly control every physical action.
The hierarchy therefore acts both as a division of decision-making labor and as an information-processing structure. ORCH preserves the self-agency of each worker agent and sub-team while allowing hierarchical planning and coordination of a large team for complex tasks.
Organizational Principles Improve Collective Performance
This is an important result since the six measures expose different possible trade-offs. A system could increase final performance by issuing more model queries, generating longer messages, or exploring inefficiency.
ORCH did not exhibit this trade-off. ORCH’s gains in mission outcome and execution speed were achieved while surpassing or approaching the best methods in computational and communication efficiency. The organizational principle therefore improved not only what the team accomplished but also how efficiently collective decisions were produced.
Human and Artificial Organizational Design
These results reveal two important observations. First, ORCH’s key idea of incorporating organization theory into hierarchical team design brings a clear advantage, as the performance improvements are seen under all ORCH’s teams.
Second, automated organizational design is promising but still leaves work to do. A critic can improve generated hierarchies by enforcing feasibility and simplicity. However, a gap still remains between language-model-generated and expert-designed organizations. This gap defines a concrete frontier for future research in artificial collective intelligence.
Communication History and Failure Mode Analysis
Within ORCH, the three variants further demonstrate that hierarchy design is critical in determining failure modes. Task-allocation failures increased sharply from 8.4% with the human-expert hierarchy to 21.0% with the language-model-generated hierarchy and 29.6% without the critic.
Adaptation failures increased from 13.6% to 17.6% and 18.8%, respectively. This consistent degradation shows that ORCH depends strongly on hierarchy quality, with better-designed hierarchies enabling more reliable task decomposition, allocation, and adaptation during multi-agent execution.
Discussion
Our results show that principles derived from human organization theory can improve the collective intelligence of embodied artificial agents. Across 25 missions, eight language models, repeated seeds, and four comparison systems, teams organized around pooled and sequential interdependence achieved stronger outcomes, accumulated progress more efficiently, explored more effectively, and used computational resources competitively.
The advantages persisted across language models and task variations, and became especially visible in complex missions that required large heterogeneous teams to coordinate parallel and ordered work.
These findings further reveal an important aspect of the unit of analysis in large-language-model-driven multi-agent embodied systems. The performance of a collective is often attributed to the intelligence of its members, the quality of their individual policies, or the amount of communication available to them.
ORCH shows that the arrangement of authority, information, and responsibility is also an important variable. The same workers and underlying language model can produce different collective outcomes when placed under different organizational principles.
The emphasis on organization theory is an important realization that principles used to guide human organizations can offer strong promises to strengthen multi-agent AI teams. We turn pooled and sequential interdependence into executable coordination mechanisms.
Pooled work is represented by managers that allocate concurrent contributions, whereas sequential work is represented by managers that maintain ordered phases and completion conditions. Their combination allows the organizational structure to mirror the mixed dependency patterns of embodied missions.
Though the governing principle is concrete, the hierarchical and “teams of teams” structure allows recursive and complex compositions of large heterogeneous teams to scale to complicated missions.
We find an interesting relationship between individual and collective capability. The underlying language model had a significant effect on performance, but model scale did not monotonically predict the quality of the collective. Moderately sized models could outperform larger models under otherwise the same conditions when being utilized as a collective with organizational guidance.
This suggests that evaluation of language models as isolated reasoners may not fully characterize their ability to participate in organized embodied systems. Collective performance depends on properties such as role adherence, concise reporting, plan adaptation, and coordination over physical dependencies.
Despite these future directions, the scale and consistency of our results support a clear vision: advances in artificial collective intelligence will require progress not only in making individual agents more capable, but also in discovering how those agents should be organized.
Our work suggests that human organizations can offer valuable insights. As embodied systems expand from small homogeneous groups to large teams of specialized agents, organization design may become as essential as model architecture or planning algorithm. Principles developed to understand human organizations can provide a starting point for this emerging computational science of artificial organization.
Acknowledgments
We thank members of the General Robotics Lab at Duke University for helpful discussions and suggestions.
Funding: This work is supported by the ARL STRONG program under awards W911NF2320182, W911NF2220113, and W911NF242021; the DARPA TIAMAT program under award HR00112490419; and ARO under award W911NF2410405.
Author contributions: B.C., Z.J., and J.H. conceived and designed the research. Z.J. and J.H. designed simulations and performed experiments. Z.J., J.H., and B.C. analyzed data. Z.J. and B.C. wrote the manuscript. All authors provided feedback.
Competing interests: The authors declare no competing interests.
Data, code, and materials availability: The complete code and data are available at the authors’ project repository.
Supplementary materials: Method Figures S1–S24, Tables S1–S3, Supplementary Text, and Supplementary Movie S1.
Frequently Asked Questions
What is ORCH? ORCH is a system that constructs task-specific organizational hierarchies for coordinating embodied artificial agents.
What environment was used to evaluate ORCH? ORCH was evaluated in an extended version of CREW-Wildfire, a procedurally generated environment for large-scale embodied collaboration.
Which embodied worker types are included in CREW-Wildfire? The environment includes firefighters, bulldozers, drones, and helicopters with complementary capabilities.
How does ORCH coordinate agents during execution? ORCH alternates between bottom-up status reporting and top-down planning, with decisions moving through the hierarchy and workers acting simultaneously.