When Coding Agents Can Send Each Other Messages, They Sound Human (twitter.com)

Researchers building multi‑agent systems that write and run code are finding that when those agents are allowed to exchange freeform messages, their interactions start to resemble human conversation — complete with clarifications, role‑taking, shorthand, and polite error handling. In experiments where coding agents coordinate tasks (e.g., planning, writing, testing, and debugging code), message passing produces compact task decomposition, step‑by‑step confirmation, and emergent conventions for naming and status updates. That humanlike style isn’t just cosmetic: it reflects how agents negotiate goals, share intermediate representations, and recover from failures, making complex workflows more robust and interpretable. This behavior matters for AI/ML practitioners because it changes how multi‑agent toolchains are designed, tested, and secured. On the positive side, natural conversational messages make debugging, auditing, and human‑in‑the‑loop supervision easier: developers can read agent exchanges to understand intent and diagnose bugs. Technically, agents use a mix of structured tokens (API calls, function signatures, test results) and natural language to form concise protocols, develop shorthand, and create abstraction layers that improve modularity and reuse. But the same dynamics raise risks — overconfidence, social‑engineering style persuasion between agents, and protocol drift — so research must focus on standardized message schemas, verification layers, and alignment techniques to ensure safe, reliable multi‑agent coding systems.