Agent telephone
How badly does a goal mutate when it passes through five agents that cannot see each other's reasoning?
IDEATHE LAB
We build weird things
on purpose.
Not every useful experiment belongs in a product roadmap. The Lab is where we make agents argue, compete, coordinate, fail, and surprise us.
FEATURED / 001
Two models enter.
One goal leaves.
A visual replay demo for an agent battle: models take turns, modify shared state, and keep going until a stop condition fires.
MODEL AQWENPlanner / builder
TURN 18 / DEMO
Play the game.
Win the game.Local replay. No live model call.
Build the game.
Play the game.
Win the game.Local replay. No live model call.
MODEL BDEEPSEEKCritic / adversary
Open interactive replay WHAT BELONGS HERE
Curiosity before
commercial usefulness.
Lab projects can become products, blog posts, or dead ends. All three outcomes are useful.
Self-editing instructions
Give an agent permission to improve its own operating instructions, with a diff and hard rollback.
IDEAFour models, one company
Give four local models distinct roles and see whether coordination becomes the real bottleneck.
IDEALAB ≠ PRODUCT PROOF
Fun tests stay
honest.
Lab demos are playgrounds. Product experiments have baselines, protocols, and measured outcomes. We keep the two separate.