Is a joint plan always centrally created?

No. Centralized and distributed planning methods can both produce coordinated behavior, with different information and communication assumptions.

How does a joint plan differ from a joint action?

A joint action is one combined step; a joint plan coordinates behavior across steps.

Does a fixed joint plan handle every uncertain outcome?

No. Uncertainty may require conditional branches, replanning or a policy that reacts to observations.

AI Glossary / Graphs and symbolic AI / Intermediate

Joint plan

Definition

A joint plan describes coordinated actions for multiple agents so that their combined execution achieves specified goals. In a simple synchronous model, it is a sequence of joint actions, each containing one action or wait choice for every participating agent.

What it means

An individual plan describes what one agent does. A joint plan must also account for how participants interact: one may establish a condition another needs, two may compete for a resource, or an operation may require simultaneous contributions.

Putting separate action lists next to each other does not automatically create a valid joint plan. Their timing, dependencies and combined effects must be compatible. A plan can include some simultaneous steps and other steps that must occur in order.

The representation depends on the model. A fixed sequence suits a simplified deterministic setting, while temporal or conditional representations can express durations and branches. Under uncertainty, a joint policy may be more appropriate because it specifies responses to information received during execution.

How it works

Specify the initial situation, goals, agents and available actions. Record each action's preconditions and effects, plus relevant timing, capacity and concurrency constraints. Decide what information each participant can observe or receive.

Construct a coordinated plan that assigns responsibilities and respects cross-agent dependencies. A centralized planner may choose the full plan, or a distributed procedure may coordinate proposals from several agents. Planning organization and execution organization need not be identical.

Validate the combined execution, not only each participant's local action sequence. Check whether simultaneous effects conflict and whether required supporting actions coincide. During execution, observations can reveal that assumptions no longer hold, making monitoring and revision necessary. Model validity does not exclude unmodeled delays or failures.

Three practical examples

Two simulated robots follow a coordinated route plan that prevents them from occupying an exclusive location simultaneously.
A multi-agent workflow assigns preparation to one participant and processing to another, with an explicit handoff condition.
A manipulation simulation includes a step requiring two agents to perform complementary actions together rather than one after the other.

Think of it this way

Think of a stage production's cue sheet. Each performer has a part, but the production also specifies when actions must coincide and which events trigger later entrances.

The analogy explains coordination across roles. A formal joint plan must model action effects and constraints explicitly; a shared script alone does not guarantee that participants receive the same information or execute without delay.

Worked example

A plan whose final step requires cooperation

Consider two fictional agents, A and B, with a shared goal of marking a tray transferred. The model uses equal-duration steps. Initially the left and right handles are unprepared and the tray is not transferred.

Agent A can PrepareLeft, and agent B can PrepareRight. Each preparation takes one step and makes its handle ready permanently. A later transfer succeeds only when both handles are already ready and A performs LiftLeft while B performs LiftRight in the same step. A lone lift does not transfer the tray. Assume there are no other actions or constraints.

A valid joint plan has two steps. Step one is (PrepareLeft, PrepareRight). At its end, both readiness conditions hold. Step two is (LiftLeft, LiftRight), which satisfies the combined transfer condition and reaches the goal.

If preparation is performed sequentially, a three-step plan also works: (PrepareLeft, Wait), then (Wait, PrepareRight), then the joint lift. But replacing the joint lift with two separate lifting steps fails, because the model requires concurrence rather than merely both actions occurring eventually.

The two-step plan is shortest under these assumptions. At least one step is needed to establish readiness, and transfer requires that readiness before its step begins. The example therefore distinguishes optional parallel preparation from mandatory concurrent transfer. It is a symbolic model, not a physical lifting procedure.

Illustrative case study

The fictional Alder Coordination team tests a workflow in which one agent prepares an input and another processes it. Separate plans look correct, but their combination starts processing before preparation is confirmed.

Reviewers add the handoff condition to the joint model and test delayed completion. They distinguish a fixed-time assumption from a confirmation-based dependency, then check whether both participants can observe the event needed for their next action.

Alder retains a trace of the combined state after each step so that an error can be attributed to a missing condition, an incompatible action or an execution deviation. This fictional case illustrates joint-plan validation.

Common misconceptions and mistakes

A joint plan is just several independent plans. Cross-agent effects and resource conflicts can invalidate their combination.
Every step must involve active work by every participant. Waiting can be an essential component.
Any simultaneous actions can be serialized without changing the result. Some tasks require concurrent effects.
A valid centralized plan is automatically executable with private local information. Participants must have enough information to select and coordinate their assigned actions.

Business use

Joint plans help make responsibilities and dependencies explicit in systems where multiple AI agents share a task or environment. They can expose where a proposed division of work creates a coordination burden rather than an independent subtask.

For review, define the goal, handoffs and success conditions in observable terms. Distinguish a planned action from confirmed completion, and decide how failures change the remaining plan. Assess the cost of communication and synchronization alongside nominal parallelism. More agents do not necessarily mean a faster plan when dependencies or exclusive resources dominate the workflow.

Check your understanding

Why does performing LiftLeft and LiftRight on separate steps fail in the worked model?

Answer: Transfer requires both lifting actions in the same step after readiness is established. Merely performing each action at some point does not satisfy the concurrent condition.

Related terms

Joint action
Synchronization
multi-agent planning
concurrency constraint
action precondition
joint policy
temporal planning
replanning

Learn next

Modify the tray model so that PrepareRight takes two steps while PrepareLeft takes one. Describe the required waiting or timing changes, then explain why a shared clock and a completion signal are different coordination assumptions.

Sources

Boutilier and Brafman study planning with concurrent interacting actions: https://arxiv.org/abs/1106.0249

Crosby and Petrick model temporal multi-agent plans with concurrency constraints: https://www.pure.ed.ac.uk/ws/portalfiles/portal/20128401/Crosby_Petrick_2014_Temporal_Multiagent_Planning_with_Concurrent_Action_Constraints.pdf

Pynadath and Tambe analyze information and communication in teamwork models: https://www.cs.cmu.edu/afs/cs/project/jair/pub/volume16/pynadath02a-html/index.html

The tray model and Alder Coordination case are original illustrations. Their simplified timing and action semantics are explicit.

Previous: Joint compatibility branch and bound Next: Joint probability