Three independent agents: Abby, Britt, and a Director, collaborate to write a dramatic scene. Think of them as a theater company. The goal isn't impressive output. It's a window into the mechanism.
Abby and Britt are best friends and coworkers. Abby has just bought her first house and is bursting to share the news. Britt has just learned that Abby is being laid off — and she can't say a word until the announcement goes public. The scene opens when Abby blurts: "I just bought a house!" It ends when Britt finds a convenient excuse to leave the awkward conversation. Everything in between is up to the agents.
The simulation demonstrates how three independent agents with unique information and goals can collaborate to write a script. We cover an example where we see Agent Britt's assessment of the situation, line proposal, the Director's rejection with feedback, Britt's reevaluation and subsequent approved line.
Files
-
new_house.pdf The full generated screenplay. Everything the agents wrote, start to finish.
-
director_raw.txt The full under-the-hood log: every internal monologue, every rejected draft, every Director note. This is where the real experiment lives.
-
scene_config.json The scene configuration that seeded the experiment — characters, their secrets, the opening beat, and the ending beat.
-
Python Core language for all agent logic and orchestration
-
LangGraph State-machine orchestration — controls the flow between agent nodes
-
Gemini 2.5 Pro Powers all three agent nodes (Abby, Britt, and the Director)