Multi-Agent Patterns: Conversational Multi-Agent
Intent
Solve tasks through open-ended natural-language conversation among agents (and optionally humans), with flexible speaker selection.
Introduction
While structured patterns like Orchestrator-Worker or Role-Based Teams enforce rigid, predefined execution workflows, many real-world problems are unstructured. In exploratory research, brainstorming sessions, and open-ended design problems, a fixed pipeline fails to accommodate creative deviations. The Conversational Multi-Agent pattern resolves this by allowing multiple agents—and optionally humans—to collaborate within a shared conversational sandbox. Coordination is achieved by reading and writing to a single, unified chat transcript, allowing agents to respond dynamically as context demands.
The Group Chat Paradigm and Shared Transcripts
Conversational multi-agent systems are typically built around three architectural components:
- Shared Transcript: A single, common conversation history that acts as the shared memory for all agents. Every message, tool call outcome, and critique is visible to all participants.
- Group Chat Manager: A coordinator component that controls speaker selection. It determines who should speak next (e.g., using a model-driven vote, round-robin, or custom heuristic).
- Human-in-the-Loop Integration: Because the medium is natural language chat, human users can easily participate as speaker nodes, providing steering commands, clarifying instructions, or auditing actions in real time.
The Context Dilution & Drift Bottleneck
The primary engineering challenge in Conversational Multi-Agent architectures is transcript growth. Because all agents write to the same transcript, the history builds up rapidly. For example, if a Planner, Researcher, and Programmer agent participate in a group chat, the raw tool printouts from the Researcher and the debug files from the Programmer are appended directly to the shared transcript. Within a few conversational turns, the context window fills with raw execution details.
This context bloat leads to two major failure modes: conversational drift (where agents lose sight of the primary goal and start chatting about minor implementation details) and repetitive looping (where agents repeat identical prompts or instructions because their attention is distracted by context noise). Consequently, this pattern requires strict context compaction techniques or is best restricted to early-stage exploratory tasks.
Trade-offs
The pattern offers high flexibility but demands careful cost and context management:
Maximal Adaptability
Ideal for unstructured, open-ended problem solving where a predefined graph or pipeline cannot be established in advance.
Seamless Human Steering
Since the medium is standard chat, human users can easily join as equal participants to audit logs, approve changes, or steer the team.
Context window bloat
Having all nodes write to a single transcript causes context windows to saturate rapidly, resulting in high token costs and attention loss.
Conversational Drift
Agents easily lose focus on the original task, descending into repetitive confirmation loops or off-topic dialogue patterns.
Known Uses
This pattern is popularized by group-chat frameworks and simulations:
- AutoGen: A pioneer framework demonstrating model-directed speaker selection in a shared group-chat manager environment [1].
- Generative Agent Societies: Interactive sandboxes (such as the virtual town in Park et al.) where agents communicate via natural-language conversation in a simulation [2].
References
-
[1]
Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., ... & Wang, C. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv preprint arXiv:2308.08155.
https://arxiv.org/abs/2308.08155 -
[2]
Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST).
https://arxiv.org/abs/2304.03442