One model per character plus a Director agent: lessons from running an AI mystery serial on DeepSeek and GLM

chappieblu · reddit · 2026-11-08

The author shares hands-on experience running a weekly mystery serial with a multi-agent setup: one agent per character plus a Director agent controlling scenes.

Setup

Actors + Director: DeepSeek V4.1 Flash — each actor sees only its own character's secrets; the Director holds the season plan, sets scene goals, and calls CUT
Narrator + Editor: GLM-5.3 — turns transcripts into chapters with noticeably better prose
A shared lorebook plus a continuity pass after each chapter

What broke and fixes

GLM's reasoning ate the whole reply budget, returning empty chapters → set reasoning effort to low
4k context made agents forget the scene mid-way → 64k mostly fixed the drift
Flash invents facts contradicting earlier chapters → lorebook + continuity pass catches most
Agents speak for each other, even faking Director lines → trim output after the first foreign speaker
The Director never called CUT on its own → an injected producer note works every time

Verifiability: the ending was locked before Chapter 4 and its SHA-256 hash published, so readers can verify the author didn't steer the story.

Original post →

One model per character plus a Director agent: lessons from running an AI mystery serial on DeepSeek and GLM

More from coding & agent