docs/vision.md

Vision

Product Thesis

Multi-Agent Land is a strange interactive theater powered by small AI agents. The user does not chat with one assistant. They drop disturbances into a living wood and watch tiny specialists argue, remember, judge, and transform the world through a shared event ledger.

Why It Fits Thousand Token Wood

• It is whimsical by default: the product surface is a playable toy, not a productivity SaaS.
• The AI is load-bearing: the agents generate the scene, critique it, and create emergent consequences.
• It showcases small models: specialist roles and compact memory make the system plausible under 32B parameters.
• It can stack prizes: polished Gradio app, agentic architecture, tiny-model mode, strong demo story, and custom UI.

Non-Goals

• Do not build a generic B2B agent platform for the first submission.
• Do not optimize for enterprise scale before the demo loop is delightful.
• Do not hide the ledger. The visible agent trace is part of the charm and credibility.

First Vertical Slice

The first slice is a single clearing in Thousand Token Wood:

1. A seed starts a run.
2. The conductor writes genesis events.
3. Three tiny agents act: scene generator, mischief critic, and pocket actor.
4. A projection renders the stage, ledger, and run stats in Gradio.
5. The visitor injects a disturbance and sees the agents react.