Running two agents with distinct personalities beats one do-everything assistant, dev says
AnimeBrian · reddit · 2026-09-21
A Reddit user describes months of experience splitting one AI assistant into two purpose-built agents ("Rick and Morty"):
- Division of labor: one handles engineering work (code, builds, debugging, terminal) with a terse, confident style that always gets human review; the other handles day-to-day tasks (research, browser, coordination) with a cautious default — verifying before claiming success, asking before irreversible actions, flagging risk.
- Temperament mattered more than the split: confident tone signals trust; hedged, double-checking tone signals to slow down and read closer — which changed how carefully the author reviewed output.
- No inter-agent chatter: everything routes through the user, which adds overhead but keeps the human in the loop.
The author invites discussion on whether deliberately running multi-agent setups with different risk tolerances beats one well-prompted assistant.
More from coding & agent
- One token is all you need: hacking LLMs into classifiers with logit_bias and max_tokens=1 — AAAzzam · 2026-09-21
- Expanso Builds a Jev-Based Agent That Reads Pod Logs and Auto-Fixes Stale Kubernetes Labels — aronchick · 2026-09-21
- Turning GPT-6 Astra Into a Motion Graphics Studio: 5 Ad Renders in 18 Minutes — PrajwalTomar_ · 2026-09-21
- Sentry CEO: if your product isn't headless, you're investing in the wrong things — zeeg · 2026-09-21
- Clairvoyance marries your AI models to your data via rich metadata layer — draginol · 2026-09-21
- Dev wires Google Astra vs Fable to play RollerCoaster Tycoon over MCP to test computer use — scaling01 · 2026-09-21