Joint Intentions

The foundational requirement for a joint intention is that all participating agents possess a mutual belief in a common goal. This is stronger than each agent individually knowing the goal; it requires that each agent believes the goal, believes that the other agents believe it, believes that the others believe that it believes it, and so on, ad infinitum. This recursive, nested belief structure ensures the goal is truly common knowledge within the team, preventing coordination failures due to private misunderstandings.

Each agent must make a social commitment not just to the goal itself, but to the joint activity of achieving it with the team. This commitment is conditional and interdependent:

• It persists as long as the mutual belief in the goal and the team's ability to achieve it persists.
• It obligates the agent to act towards the goal while appropriately coordinating with others.
• It creates an obligation to inform teammates if the agent can no longer participate, triggering a re-evaluation of the joint commitment. This transforms a collection of individual intentions into a unified, accountable collective.

Agents operating under a joint intention must be mutually responsive to the actions and intentions of their teammates. This characteristic ensures the collaboration is dynamic and adaptive. It involves:

• Monitoring teammate actions and the environment.
• Adjusting one's own actions to mesh with the team's progress.
• Supporting teammates who are struggling with their sub-tasks.
• Avoiding actions that would undermine the collective effort. This responsiveness is what differentiates a team from a set of agents executing pre-programmed scripts in parallel.

A joint intention does not dissolve simply because one agent decides to stop. It is persistent until a specific termination condition becomes mutual belief among the team. Termination occurs when the team mutually believes either:

1. The goal has been achieved.
2. The goal is no longer achievable.
3. The goal is no longer relevant. This persistence condition prevents premature abandonment and requires a joint exit ritual, such as communication to establish the mutual belief that the endeavor is over, ensuring a clean and coordinated conclusion to the collaborative state.

A joint intention is not equivalent to the mere conjunction of individual intentions toward the same goal ("I intend that we J" vs. "We intend that we J"). Key distinctions include:

• Interdependence: My commitment is contingent on your continued participation.
• Obligations: I acquire social obligations to you (e.g., to inform you of problems).
• Robustness: The joint activity can often survive the failure of an individual plan, as the team can replan.
• Exit Conditions: I cannot unilaterally drop the intention without violating a social commitment to the team. This framework is crucial for designing agents that can engage in robust, human-like teamwork.

The theory is formalized using modal logics of belief, intention, and action (e.g., BDI logics). In practical multi-agent systems, implementations approximate these ideal conditions using:

• Shared Plan Representations: Using structures like Hierarchical Task Networks (HTN) to represent the common goal and sub-task structure.
• Communication Protocols: Employing Agent Communication Languages (ACL) with speech acts like inform, request, and confirm to establish and maintain mutual beliefs.
• Teamwork Models: Frameworks like STEAM (Shell for TEAMwork) operationalize joint intentions by defining team plans with explicit commitment and synchronization points, translating theory into executable agent code.

Social Commitments are normative constructs that create explicit, directed obligations between agents. They formalize that a debtor agent is committed to a creditor agent to bring about a specific condition. This provides a crucial mechanism for establishing accountability and trust in open systems, forming the contractual backbone that makes Joint Intentions actionable.

• Key Difference: While Joint Intentions represent a team's mutual commitment to a goal, Social Commitments define the bilateral obligations between individual agents that underpin that collective state.
• Example: In a delivery task, an agent might form a social commitment to another agent to "hand off package X at location Y by time Z," contributing to the joint intention of completing the delivery.

The Belief-Desire-Intention (BDI) architecture is a software model for intelligent agents based on the philosophical theory of practical reasoning. It provides the individual cognitive framework upon which Joint Intentions are built.

• Beliefs: The agent's knowledge about the world (including beliefs about other agents' beliefs).
• Desires: The agent's goals or motivational state.
• Intentions: Desires the agent has committed to pursuing; they are persistent and constrain future reasoning.

Joint Intentions theory extends the BDI model to the multi-agent case, requiring not just individual intentions but mutual beliefs about those intentions and their collective commitment.

An Agent Communication Language (ACL) is a formal language with precisely defined syntax, semantics, and pragmatics that enables autonomous agents to exchange information and knowledge. It is the practical medium through which agents establish and maintain Joint Intentions.

• Foundation: Based on Speech Act Theory, where messages are actions (e.g., inform, request, cfp - call for proposal, propose).
• Standard: The FIPA ACL is a prominent standard that defines communicative acts and associated protocols.
• Role in Joint Intentions: To form a joint intention, agents must communicate to establish the mutual belief that they have a shared goal. An ACL provides the verifiable, unambiguous messages needed for this process (e.g., "I commit to goal G if you do").

Partial Global Planning (PGP) is a coordination approach where agents exchange and merge their local plans to identify and resolve potential interactions, conflicts, or opportunities for cooperation. It is a procedural methodology that can be used to achieve the state described by a Joint Intention.

• Mechanism: Agents develop local plans, communicate key elements (goals, expected results, resource needs), and form a "partial global plan"—a non-centralized, coordinated view of activities.
• Relation to Joint Intentions: PGP is a process for planning how to fulfill a joint intention. The joint intention provides the persistent collective commitment; PGP provides a dynamic, negotiated plan to realize it, which can be revised as beliefs change.

Electronic Institutions are computational frameworks that define the norms, rules, and structured interaction spaces governing the behavior of autonomous agents to ensure orderly and goal-directed societal interactions. They provide the organizational context in which Joint Intentions are formed and executed.

• Components: Define roles, interaction protocols (dialogues), norms (obligations, permissions, prohibitions), and virtual spaces (e.g., a bidding room).
• Function: They create a predictable, regulated environment that reduces uncertainty, making it safer for agents to form commitments and joint intentions. They enforce consequences for broken social commitments.
• Analogy: If Joint Intentions are a "contract" between agents, an Electronic Institution is the "legal system" and "courtroom" that defines valid contracts and adjudicates disputes.

SharedPlans Theory is a formal theory of collaborative planning that deeply complements and extends Joint Intentions. It provides a richer model for how a group constructs, monitors, and executes a plan together.

• Core Focus: Details the mental states (knowledge, intentions) required for an agent to participate in a shared plan, including intentions toward the actions of other agents.
• Key Constructs: Introduces recipe (a known method for achieving a goal) and potential for choice (which agent will perform a sub-action).
• Integration: While Joint Intentions theory establishes the commitment to a shared goal, SharedPlans theory elaborates the commitment to a shared process for achieving it. They are often used together to model full collaborative problem-solving.