Barbara Grosz and Sarit Kraus: SharedPlans and the Computation of Collaboration
Barbara Grosz and Sarit Kraus developed SharedPlans into a detailed computational theory of collaborative action, including partial knowledge, commitment and complex group activity.
Collaboration is not the same thing as simultaneous action
Two agents can perform actions at the same time without truly collaborating. Collaboration requires relationships among goals, plans, responsibilities and knowledge. Barbara Grosz and Sarit Kraus’s SharedPlans work sought to formalize those relationships for complex group activity.[1]
Their 1996 Artificial Intelligence paper extended an earlier SharedPlans model to address multi-agent actions, commitment to joint activity and situations in which participants know only part of the overall plan.
SharedPlans avoided requiring one agent to control another’s intentions
A central design choice was to model collaboration without requiring one agent to possess intentions toward the internal acts of another agent. The framework instead represents how individual intentions, beliefs and commitments relate to the shared activity.[2]
That makes collaboration compatible with autonomy: participants can remain distinct agents while still coordinating toward a joint objective.
The group plan can be incomplete
Agents do not need complete knowledge of how every subtask will be performed. SharedPlans explicitly supports partial plans and partial knowledge, which is critical when expertise and information are distributed.[1]
Contracting out actions fits the model
The expanded framework allows actions to be delegated or contracted, showing how a collaborative plan can incorporate work performed by agents outside the immediate decomposition of the activity.
Commitment gives collaboration persistence
A collaboration cannot be modeled only as a momentary coincidence of goals. Participants must have reasons to continue supporting the joint activity and to communicate when relevant conditions change. The 1996 model addressed the need for agents to commit to shared activity.[2]
This connects formal planning with the practical reality that team members depend on one another’s continued participation.
SharedPlans linked dialogue research to multi-agent systems
Grosz’s earlier work on discourse had already treated conversation as structured, intentional activity. SharedPlans extended related ideas into computational collaboration, where communication supports coordinated action rather than merely exchanging isolated messages.[3]
Her Harvard research record explicitly identifies collaborative planning and multi-agent collaboration as major contributions.[5]
Communication has a reason inside the joint plan
In a collaborative system, agents should communicate because a change matters to the shared activity—not simply because new information exists. This gives information sharing a task-sensitive foundation.
The theory scaled from pairs to complex group action
The Grosz-Kraus paper broadened SharedPlans beyond simple activities that decompose directly into single-agent steps. It accommodates complex actions involving groups and nested plans, making it more suitable for realistic teamwork.[1]
Bar-Ilan’s publication record documents the article’s scope and its relationship to earlier formulations of SharedPlans.[4]
Collaboration became an AI systems challenge in its own right
In her AAAI presidential address, Grosz argued that collaboration must be designed into intelligent systems rather than patched on afterward.[4] The argument positioned collaborative capability as a foundational AI problem spanning planning, language, reasoning and human-computer interaction.
This widened the meaning of multi-agent research: the aim was not merely to coordinate machine components but to create systems capable of productive partnership.
SharedPlans influenced later teamwork and human-agent research
The model became a reference point for later work on agent teamwork, intention reconciliation, task allocation and collaborative assistants. Grosz’s publication list shows a continuing research program around collaboration, social reasoning and information sharing.[3]
The 1996 paper later received an influential-paper award, reflecting its long-term role in the field.[5]
The key abstraction is mutual support for a joint activity
SharedPlans helps explain why a team is more than a set of independent optimizers. Participants reason not only about their own next action but about what must be true for the joint activity to succeed.
Why Grosz and Kraus belong in multi-agent history
Their work gave collaboration a precise computational structure while preserving autonomy, partial knowledge and complex decomposition. It connected formal models with the real requirements of teamwork: commitment, communication and adaptation.[1][2]
Modern agent systems that assign roles and share context still face the same conceptual challenge: when does a collection of agents become a collaborating team? SharedPlans remains one of the field’s foundational answers.
Works Cited
- 01Artificial Intelligence — Collaborative Plans for Complex Group Action sciencedirect.com
- 02Harvard DASH — Collaborative Plans for Complex Group Action dash.harvard.edu
- 03Barbara J. Grosz — Publications grosz.seas.harvard.edu
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead