Blog article Artificial intelligence
Before the prompt: designing a conversation’s architecture
When several agents participate, who can reply to whom matters. A guide to roles, turns, memory and evaluation before automating a discussion.
Vinicius Covas3 min read
Conversation also has a design
Imagine a working group where everyone takes a turn but nobody can directly reply to another person. Now allow replies and add a moderator who chooses which question to revisit. Participants and topic remain the same, but the possibilities for interaction have changed. This hypothetical example helps us think about multi-agent systems: listing models and assigning names is insufficient. We need to describe the communication setting we construct for them. Its rules may enable comparison of arguments or turn the exercise into a sequence of monologues.
A research question
The paper presented at the MALGAI workshop at ICLR 2026 studies communication modalities and interaction networks among agents. The full reference and publication status appear below. Here I want to develop a design question suggested by that research agenda: what kind of conversation should be possible? A system exploring objections needs opportunities for disagreement; one organizing information needs explicit selection criteria. Giving both tasks the same structure may hide important differences before we even evaluate an answer.
Four decisions before running the system
First, I would define each role’s purpose without confusing a label with actual competence. Next, I would specify which information each agent can consult and which messages it receives. The third decision concerns replies: who can be answered, and how often? Finally, I would define how discussion ends and who reviews the result. A turn limit prevents endless conversation but does not demonstrate that a sufficient conclusion has been reached. The stopping condition should therefore remain separate from the criterion used to judge quality.
Compare without changing everything at once
To explore a decision, I would keep the task and sources fixed while changing one interaction rule. For example, compare an independent round with a round allowing replies to an objection. Save messages and record execution conditions so the process can be reviewed. I would look beyond a more convincing conclusion to ask which evidence disappeared, which disagreement went unanswered and which unsupported claims circulated. This is a proposal for critical examination, not a recipe guaranteeing better results in every model or setting.
Responsibility remains a design decision
A group of agents can produce a unified answer without reaching well-supported agreement. Preserve the path to the synthesis and allow a person to review its support. A real application also needs to specify which actions require authorization and how to stop a mistaken sequence. I propose starting with these questions and choosing tools afterward. Designing a conversation means organizing opportunities to speak, reply and correct; evaluating its outcome requires checking whether those opportunities served the intended purpose.
Communication researcher and professor. I study how humans and machines build meaning and relationships.
About the author