Multi-agent demos you can actually reproduce

Description and use

Separating personas, run logic, and UI in the Séance template yields deterministic, testable multi‑agent demos. Property‑based tests uncovered and fixed a scoring bug, and the template supports multiple domains without engine changes.

Problem

Multi-agent demos often entangle persona definition, run logic, and UI rendering, making testing, replaying, and updates difficult.

Outcome

By separating personas, run logic, and UI, the author achieved deterministic, testable multi-agent demos. Property-based testing uncovered a scoring bug, which was fixed by adjusting the formula. The template supports multiple domains without engine changes.

Workflow

• Define personas in YAML files.
• Implement runRoom({artifact, personas, seed}) as a pure function that returns a Transcript.
• Render the Transcript in the UI.
• Use a seed to control wording without randomness or I/O.
• Run property-based tests with fast-check to validate invariants.
• Fix scoring bug by using a capacity independent of findings.
• Deploy the template with two demo apps without modifying the engine.
• Re-run demos with a new seed to see voice changes without changing evidence.

Evidence

Séance splits them: personas are YAML, the run is a pure function, the UI takes a transcript.

Source

No Math.random(), no Date.now(), no crypto.randomUUID() inside seance-core.

Source

The score could go down when you pointed at a new risk.

Source