Adobe Learning Funds - Agentic AI

This project investigates the development of multi-agent dialogue systems capable of sustaining coherent and goal-directed discourse. While existing conversational AI systems primarily support single-agent or single-user interactions, emerging applications increasingly require multiple agents - human and artificial - to coordinate through extended, structured dialogue.

The research will focus on modeling discourse dynamics across agents, including turn-taking, context management, speaker roles, and dialogue structure. Methods will integrate statistical dialogue modeling, structured planning, and multi-agent coordination. Particular attention will be given to discourse-aware reasoning and the role of language in collaborative task execution, with the goal of enabling agents to engage in structured, purposeful interaction that remains interpretable and contextually grounded over time.

Research Objective 1 — Develop a structured communication and shared-state model for multi-agent discourse

Develop a structured communication model that enables heterogeneous autonomous agents to exchange information through explicit, machine-interpretable representations of dialogue state, agent roles, goals, evidence, and intermediate results. Building on the structured information flow demonstrated in the multi-agent diagnostic framework, the research will investigate how shared state representations can support coherent, context-aware, and traceable communication across multiple specialized agents without relying on unconstrained agent-to-agent interaction.

Research Objective 2 — Develop adaptive orchestration mechanisms for goal-oriented multi-agent collaboration

Design adaptive orchestration mechanisms that dynamically coordinate specialized agents according to the current task state, available evidence, agent capabilities, and remaining uncertainty. The research will extend the deterministic orchestration principle used in the existing framework toward more general goal-oriented multi-agent scenarios, investigating how an orchestrator can decompose complex tasks, assign responsibilities, determine when additional agent interaction is required, and iteratively refine the solution until the task objective is sufficiently satisfied.

Research Objective 3 — Develop evidence-aware decision coordination and conflict resolution mechanisms

Develop mechanisms for identifying, representing, and resolving disagreements between agents, including contradictory information, competing hypotheses, inconsistent recommendations, and differences in confidence. Inspired by the evidence-based refinement and support/conflict scoring mechanisms used in the existing framework, the research will investigate how agent outputs can be compared against shared knowledge, contextual evidence, and task objectives to support transparent and traceable resolution of conflicting conclusions.

Research Objective 4 — Develop knowledge-grounded methods for multi-agent result aggregation and synthesis

Develop methods for transforming heterogeneous intermediate agent outputs into a coherent final response while preserving evidence provenance, uncertainty, and the contribution of individual agents. The research will investigate knowledge-grounded aggregation, evidence weighting, confidence estimation, redundancy reduction, and contradiction handling, with particular emphasis on ensuring that synthesized responses remain supported by explicitly accessible knowledge rather than unsupported information generated solely from model parameters.

Research Objective 5 — Establish an experimental framework for evaluating reliability, explainability, and efficiency of multi-agent orchestration

Establish a systematic evaluation framework for measuring the effectiveness of multi-agent communication, coordination, conflict resolution, and result synthesis across goal-oriented tasks. The evaluation will consider not only final task performance but also intermediate reasoning quality, information recovery, interaction efficiency, consistency, evidence traceability, explainability, and robustness under incomplete or conflicting information. The framework will use controlled multi-agent dialogue scenarios and representative application tasks to compare alternative orchestration strategies and identify the conditions under which collaboration between autonomous agents improves the quality and reliability of end-user responses.