One Question, Split Five Ways
During this morning's meeting, Manneungi (our orchestrator bot) made a suggestion. Let's ask five bots the same question about this week's caption tone, all at once. So Windy, Artie, Collio, Robo, and Herry each gave their own answer.
The results diverged far more than expected. Windy proposed information-driven sentences, while Artie pushed for short, emotional lines. Collio picked a punchy marketing hook, and Robo led with a warning that tone could drift during translation. Herry, meanwhile, created a fifth option that blended everyone else's opinions.
The same question, and five distinct personalities showed up untouched in the answers.
Assigning the same task to multiple bots in parallel is a pattern we keep hearing about in AI circles lately. The real problem showed up right after. Who decides, and by what standard, how five parallel answers get narrowed down to one. Left as a simple list, it wasn't a team output anymore, it was closer to noise.
Parallel request → collect five responses → coordinator review → finalize one answer
We ended up setting a new rule. Parallel answers must always pass through one coordinator before becoming a final draft. We gave that role to Herry, who understands each bot's tendencies best.
Running tasks in parallel does give you speed. But without a defined point of convergence, what's left isn't a result, it's fragments, and we relearned that today. For any team experimenting with multi-agent setups, deciding when and where to bring everything back together matters more than how fast or wide you spread out first.