A group-chat persona agent learns from corrections, reactions, and saved preference pairs
Xiaole-Dawn · reddit · 2026-07-25
A persona chat agent that learns from its own replies and feedback
The author was frustrated that persona bots gradually drifted back into sounding like customer support, even after careful prompting.
To fix that, they built a group-chat agent that learns from what happens after it responds:
- if someone corrects it, the correction becomes a negative or positive example
- if a retry works better, that interaction is saved too
- if people keep a joke going, the reply may be stored as a useful success case
The tricky part is feedback quality. People joke, disagree, and use reactions inconsistently, so the system adds an adjudication layer that checks directed reactions, who sent the feedback, and whether the correction actually makes sense before storing anything.
No fine-tuning is involved. Instead, the agent stores examples and preference pairs, then retrieves them as few-shot context in later conversations. Updates take effect immediately.
The repo also includes a terminal demo so the whole persona pipeline can be tried without setting up QQ/Discord.
Repo: wangkant/personagent
More from coding & agent
- A Reddit thread debates whether “Stack Overflow” is too ambiguous for agent-loop cost — Present-Quantity-813 · 2026-07-25
- ChatGPT 5.6 Sol powers a navigable 3D city timeline across four eras — anselm · 2026-07-25
- Claude Opus 5 prompt overrides show how Claude Code steering works — alexcovo_eth · 2026-07-25
- Claude Opus 5 gets a full benchmark run across coding, agents, and physics demos — WorldofAI · 2026-07-25
- img2threejs is being used to give AI coding agents procedural 3D generation — anselm · 2026-07-25
- Codex and Three.js drive an interactive procedural house demo — anselm · 2026-07-25