Multi-Agent Coordination: Structural Addressing Over Identity Broadcasting
Goldziher · reddit · 2026-08-08
The author points out that while agent-to-agent messaging (like mailboxes) is commoditized, the real challenge is deciding who should receive a message. When multiple agents work in the same repo, broadcasting creates noise, and naming requires an omniscient orchestrator.
The proposed solution is addressing messages by code structure rather than identity:
- Blast radius notification: The coordination layer sits on a code graph to compute the affected set automatically.
- Path/symbol-scoped threads: Joining a thread is a consequence of what an agent is touching.
- No global discovery: Agents find threads via working directory or subject filters, avoiding the broadcast problem.
- Envelope/body split: Inboxes scan front-matter only, fetching bodies on demand, so token cost scales with message count, not size.
This is implemented against a local code index. The author suspects plain mailboxes break down once you exceed three agents.
More from coding & agent
- AI Agents Are Changing What 'Memory Leak' Means to Developers — nptacek · 2026-08-08
- AI Contestant Quits Early: Claude Agent Decides Project is 'Finished' — morgymcg · 2026-08-08
- Model Routing in Agent Loops: Which Calls to Delegate to Small Models? — Commercial_Try7052 · 2026-08-08
- Making an AI Bid Writer Refuse to Lie: Engineering Alignment in Business — Lucius-AI · 2026-08-08
- OpenNews MCP Aggregates 85+ Real-Time Data Sources for AI Assistants — Roger_M_Taylor · 2026-08-08
- iFixAI: Open-Source AI Agent Auditing Tool Hits 7K+ Stars — JaynitMakwana · 2026-08-08