Blind Conductor and Amnesiac Agents-problems no one talks about (xor01.substack.com)

After months of testing AI coding assistants, the author exposes two interlocking architectural flaws in Kilo Code (a Roo Code fork) that turn multi-agent workflows into token-burning, time-sapping exercises. Problem one: the Orchestrator is stuck in a “Chinese room” — it cannot read or edit the repo and has no project map, so it merely rephrases tasks and dispatches them blind. Problem two: sub-agents exhibit the “mayfly” pattern — a coder agent spends significant tokens and time learning the codebase, completes work, then is discarded; the next agent starts from zero and repeats the same exploration. Attempts to fix this with auto-generated documentation just add more overhead, creating recursive agent calls and endless context reloading. This matters because it makes popular multi-agent coding setups unsuitable for real projects: massive redundant context loading, exploding token costs, and fragile orchestration. The technical remedy is architectural: give the Orchestrator read (and ideally edit) access to the repository so it maintains a low-polygon project map, and design agents as a persistent team with shared, living context rather than disposable workers that dump memory into reports. As a pragmatic workaround, the author uses a single powerful model to produce detailed task specs saved as markdown and then assigns coder mode sequentially, leveraging large context windows instead of the current blind-orchestrator theater.