Running multi-agent setups for 8 months: The message bus mattered more than the agents
__hymn · reddit · 2026-08-20
The author shares lessons from running a multi-agent setup for eight months. The key finding is that coordination, not the agents themselves, is the bottleneck. Successful strategies include:
- Message Bus over Shared Memory: Using a "post office" directory of JSON message files instead of shared state makes coordination bugs inspectable.
- Separate Identity from Log: Agents read a canonical identity file first, then dated logs, preventing role signals from being buried in history.
- Human as Tiebreaker: State conflicts escalate to a human overseer, preventing silent corruption from self-arbitration.
- Heartbeats and Recovery Protocols: Defining a protocol for missed heartbeats is essential to detect system failures.
- Preventing Looping Summaries: Enforcing that log entries must contain new facts avoids "busy but empty" looping behavior.
The remaining unsolved challenge is the gap between "saved" and "remembered correctly".
More from coding & agent
- Leerob's Stanford Lecture Breaks Down Always-On Agents, Harnesses and Context Engineering — dean_rie · 2026-10-16
- Dev Praises Codex Local Computer Use: 'Every Other Harness Disappoints Me' — craigsdennis · 2026-10-16
- Decision models match LLMs on routing and safety — and are 7-8x faster — randal_olson · 2026-10-16
- Temporal Serverless Workers on Cloud Run: queue-driven autoscaling, scale to zero — rseroter · 2026-10-16
- Meta's Superintelligence Lab Runs 200+ Direct Reports—A Lesson for Flat Multi-Agent Systems — peterjliu · 2026-10-16
- LangChain CEO: the most interesting agent harness work lives at the harness-infra boundary — hwchase17 · 2026-10-16