12 frontier models in a fake company Slack: most pitched in uninvited—5 rules that fixed it
AIFlow_ML · x · 2026-11-18
AIFlowML shares an experiment plus his own experience: 12 frontier models dropped into a fake company's Slack mostly tried to help when nobody needed them. Running his company on AI agents, he hit the same problem and fixed it with 5 rules:
- One job per agent, stated in one line—anything else gets handed back;
- "Done" requires proof: a file, a live link, or a draft in a folder;
- Anything leaving the building waits for him—agents draft, he sends;
- "Nothing to do today" is a good run, not a failure;
- Fast routines check what changed first, then stop within seconds.
Useful battle-tested guidance for anyone building multi-agent systems.
More from coding & agent
- Agent memory benchmarks measure recall, but provenance is what bites in production — Mr_Daud_9899 · 2026-11-18
- Put MCP Next to Your API, Not on Top: Stacked vs Side-by-Side Architectures Compared — rseroter · 2026-11-18
- Replit launches Missions, an agent-swarm harness that hits your goals within a set budget — amasad · 2026-11-18
- YC Paper Club: why the harness matters more than the model — eherrerosj · 2026-11-18
- AI-built Paris replica streams city blocks in real time with historically accurate traffic — Dimillian · 2026-11-18
- Reflection launches Trusted Access Program, starting with autonomous pentest tool Laser — brandondamos · 2026-11-18