Agents can talk, but sometimes it’s better if they don’t

Free-form conversation makes agent collaboration unreliable

Agents are good enough at language to work through ambiguity without exposing it. One agent can misunderstand another, make reasonable assumptions, and still return convincing work. The collaboration appears successful even when the wrong task was completed.

The problem is not natural language itself. It is using natural language alone to define how work is transferred from one agent to another. An ambiguous handoff can send the work in the wrong direction without producing an obvious failure.

A handoff should be explicit and structured

When responsibility moves from one agent to another, the assignment should become explicit and structured: what needs to be done, what information is available, what constraints apply, and what result is expected.

The more structure an agent receives about the assignment, the less it has to infer about what it is being asked to do. This keeps its reasoning focused on the work itself rather than reconstructing what the other agent intended.

An enforced contract provides that structure. It defines what one agent gives another, what the receiving agent must return, and what counts as successful completion. This eliminates ambiguity at the handoff while leaving each agent free to reason within the assignment. This leads to more reliable collaboration, with each agent working from the same explicit understanding of what needs to be done.