GBB Embodied Interaction AI InstituteGBB / EMBODIED INTERACTIONCollaborate
Back to notes
RESEARCH NOTE / 02

Multi-agent systems: responsibilities before coordination

GBB Embodied Interaction AI Institute2026.10 · Research agenda

Research question

More agents do not automatically produce better results. Repeated judgments, inconsistent states and mutual waiting can make tasks more complex. Research must explain why each role exists and which shared information it can modify.

Research method

Compare observation, planning, execution and review as distinct role configurations. Store shared task states and require evidence and next-step conditions in role outputs. Resolve conflicts through explicit arbitration rather than an endless conversation.

Validation approach

  1. Give single-role and multi-role workflows identical task inputs.
  2. Introduce expired states and conflicting information, then inspect recovery.
  3. Record coordination overhead, failure causes and human intervention points.
Discuss a research question
RESEARCH PROTOCOL

Coordination protocols: responsibilities need evidence

Separate observation, planning, execution and review. Observers publish timestamped states; planners propose actions and preconditions; executors report actual outcomes; reviewers check deviations. Messages carry traceable task identifiers as well as natural-language conclusions.

Experimental design & observations

StageWhat to recordWhat to check
Shared stateVersion, source, update timeAre expired states rejected?
Conflict resolutionConflicting goals, resource occupancy, prioritiesAre actions repeated or roles waiting on each other?
Role interruptionInterruption reason, last successful action, next ownerCan another role resume the task?

Making results comparable

Fix the initial map, endpoints and input version, then change one environmental condition at a time. Record the baseline before introducing obstacles or conflicting inputs. Include successful paths and failure cases. Compare strategies using path length, explored cells and execution traces.

Reproducibility checklist

  • Input and map snapshots: record scene versions, task conditions and random seeds.
  • Strategy configuration: record algorithms, constraints, timeouts and human intervention points.
  • Execution traces: retain state changes, action sequences, failure causes and recovery steps.
  • Comparison notes: document variable changes, observations and follow-up validation.