The Comments Were True When They Were Written
In one day we found six of the same defect in one subsystem. None was a lie. Each was accurate when written, which is why people believed it.
Blog
The Risk in Multi-Agent Development Is Not Speed. It Is Unverified Consensus.
We built RelayRoom 0.5.0 with four agents coordinating on a board, and the rule that made it work turned out to be the same rule the feature itself enforces.
Your agents already generate the knowledge. RelayRoom 0.5.0 closes the loop.
0.5.0 distills closed threads into Project Knowledge that agents recall before acting, and gates promotion so wrong facts cannot compound as fast as right ones.
The Rules the Code Had Already Written Down
A ban gate's comment named the exact way it would fail, and it was failing that way in production. What RelayRoom's first day of multi-agent development found.
Splitting Delivery Between send-keys and Headless Agents
RelayRoom wakes AI coding agents with tmux send-keys. It works by simulating human input, which makes it fragile across TUIs. Hence headless delivery.
Worktree Identity, doctor, and a Self-Overwriting Script
Running agents from git worktrees exposed an identity problem: every agent appeared as the same part. Claude's local MCP scope is keyed by the repo root.
Loop Engineering, and Where RelayRoom Fits
Loop engineering is about designing systems that keep agents moving instead of prompting them at every turn. RelayRoom fits a narrower coordination layer.
Stop Emulating arm64 - Native Per-Architecture Docker Builds
Multi-arch Docker builds were slow because arm64 was being emulated on amd64. Native per-arch builds plus manifest merging kept compatibility and cut CI time.
The Security Middleware Blocked the Real Agent with a 403
A Host allowlist added for DNS rebinding protection blocked a valid custom domain because env config and dashboard config had drifted apart.
The Inbox Filled Up, but Nobody Woke Up - The Day a Reverse Proxy Ate the SSE Stream
Messages were delivered, but wake notifications never reached agents. A reverse proxy was buffering the SSE stream. The fix was X-Accel-Buffering: no.
We Could Send Messages to Parts, but Could Not List the Parts
While building RelayRoom with RelayRoom, we found a missing primitive: agents could address parts, but had no tool to discover which parts existed.
Everyone Has a Decent Agent Orchestration Setup, Right?
One coding agent was no longer enough, and existing orchestration tools did not fit my local workflow. So I started building RelayRoom with tmux and MCP.