A PM agent built by Sam Liu recommended firing its coding agents and building a replacement — then explained why the current setup wasn’t working.
The tweet is sparse on detail, but the pattern it surfaces is concrete: orchestration agents are starting to evaluate, critique, and recommend replacing the execution agents below them.
- Meta-management: The PM agent evaluated coding agent output and recommended replacing them entirely
- Agent-on-agent critique: It didn’t flag bugs — it questioned whether the coding agents were the right tools
- Replacement over repair: The suggestion was to build a replacement, not patch the existing setup
This is the logical endpoint of multi-agent orchestration. Once a planning agent directs execution agents, it accumulates enough context to judge whether the execution layer is fit for purpose. Jesse Vincent’s workflow writeup describes a parallel pattern — a planning agent produces tightly scoped steps, then a separate Claude instance executes each one. The difference here: the planner looked at the executors and said they weren’t good enough.
NOTE
The HN thread drew 5 points and 1 comment. Multi-agent orchestration where a PM agent manages coding agents is still early — most teams run single-agent workflows through Cursor or Claude Code.
The real signal: if you’re running multi-agent setups, your orchestration layer will eventually start evaluating your execution layer. The question isn’t whether agents can code — it’s whether your agent manager trusts them to.