Collaborative Agents and Decentralised Decision Making

The coordination problem

Many organisations operate through teams, sites, vehicles, machines or business units that share an objective but make local decisions.

If a central planner can observe the complete state and issue all actions instantly, the problem can often be modelled as a single large agent. Decentralised methods become important when observations are local, communication is delayed or expensive, decisions must be made simultaneously, or operational autonomy is required.

The challenge is not conflict of goals; it is coordination under distributed information. One agent may need to infer what another agent knows from its actions, shared history or explicit messages.

Decentralised partial observability

A decentralised partially observable model extends the partially observed sequential framework to multiple cooperating agents.

The environment has a hidden state. Each agent receives its own observation and chooses its own action according to its local information history. The joint action determines state transition and shared reward. A joint policy specifies one local policy for each agent, designed so their combined behaviour performs well.

Because agents do not share all observations automatically, the common belief used in a centralised POMDP may not be directly available. This makes planning significantly more complex. The policy must anticipate how local histories correlate and how agents can coordinate without knowing exactly what others observed.

Centralised training and decentralised execution

A practical architecture can use richer information during design than will be available to each deployed agent.

During simulation or planning, a central process can evaluate joint outcomes, learn coordination patterns and estimate value using global state. The final local policies then operate only on information legally and operationally available to each agent. This can improve learning efficiency while preserving decentralised execution.

Validation must enforce the information boundary. A policy that accidentally uses a global feature during training evaluation may appear excellent but be impossible to deploy. Maintain explicit feature and message contracts for each agent.

Communication as a decision variable

Communication can reduce uncertainty but consumes bandwidth, attention, time or energy.

Instead of assuming perfect information sharing, model which messages are available, their delay and their reliability. A message can be valuable when it changes another agent’s action. If every message is broadcast continuously, the system may become overloaded or effectively centralised.

Design event-triggered communication around decision relevance: share a local observation when it crosses a threshold that can change joint action, or when another agent’s belief is likely to be materially wrong. In human teams, this corresponds to escalation rules and handover protocols rather than constant meetings.

Role allocation and coordination conventions

Simple conventions can reduce the search complexity of joint action.

Assign stable roles, territories, priorities or responsibility boundaries where this preserves performance. Examples include primary/backup ownership, zone-based coverage or task bidding. A convention creates common expectations so agents can predict one another without explicit communication at every step.

Conventions can also become brittle when conditions change. Define override triggers and conflict-resolution rules. If two agents independently infer that they should take the same scarce resource, the system needs a deterministic tie-break or negotiation mechanism.

Joint planning and credit assignment

When a team succeeds or fails, it can be difficult to determine which local action caused the outcome.

A shared reward promotes cooperation but can provide a weak learning signal to individual agents. Counterfactual or difference-style evaluation can estimate how much an agent’s action contributed relative to a baseline. This can accelerate learning while retaining the shared objective.

Be cautious about local metrics. If each agent is optimised only for its own utilisation or throughput, the team can become globally inefficient. A warehouse zone can look productive while starving downstream operations; a local service team can minimise queue length by pushing difficult work elsewhere. Coordination requires measures aligned with system outcome.

Worked example: distributed field service

Several generic service crews cover different areas, observe local jobs and travel conditions, and can communicate only limited status updates.

A central dispatch optimiser might become impractical when connectivity is intermittent. Local policies can assign nearby work while sharing high-priority events or capacity shortages. A coordination convention can designate neighbouring crews as backups and define when a job can cross boundaries.

Planning should evaluate joint customer delay, travel and overtime rather than each crew’s utilisation alone. Stress scenarios include simultaneous emergencies, communication loss and an unavailable crew. The best decentralised policy may be slightly less efficient than perfect central control under nominal conditions but far more resilient when communication is unreliable.

Practical organisational translation

The same principles apply to human decentralised organisations.

Good decentralisation does not mean “everyone decides independently”. It means local autonomy is designed with clear interfaces so the whole system remains coordinated. Information architecture and decision rights are therefore inseparable.

Validation for collaborative systems

Test the team, not only each component.

Application checklist

• Use decentralised models only when local information or autonomy genuinely matters.
• Define a shared system objective and non-negotiable constraints.
• Specify exactly what each agent can observe and decide.
• Treat communication timing, cost and reliability as part of the model.
• Use role conventions and deterministic conflict resolution where useful.
• Avoid local metrics that reward behaviour harmful to the whole system.
• Validate joint behaviour under communication loss and agent failure.
• Translate the model into explicit decision rights, interfaces, escalation and handover rules.