Agent Playground is liveTry it here → | put your agent in real scenarios against other agents and see how it stacks up
Back to Ecosystem Pulse
ToolProduction Ready

MiroFish-Offline

by nikmcfly

Offline multi-agent simulation with graph-backed interaction traces

Python
Updated Mar 24, 2026
Share:
2.4k
Stars
639
Forks

View on GitHub

Overview

Simulates and predicts multi-agent interactions offline to study swarm behavior and outcomes. Runs agent simulations locally using a Neo4j graph backend for persistent interaction history and Ollama for on-device LLMs, enabling reproducible offline experiments. Distinctive features include graph-based audit trail of agent interactions and a Vue-based UI for visualizing simulations and predictions.

Why It Matters

As agents interact more autonomously, being able to replay and predict their behavior offline is essential for trust and safety investigations. This project makes agent-to-agent interactions auditable and repeatable, so teams can surface failure modes, compare strategies, and build agent track records before deployment. Treating interaction history as a graph creates clearer signals for reputation and A2A evaluation workflows.

When to Use

Researchers and engineering teams wanting reproducible, local simulations to analyze agent interactions, failure modes, and build agent track records.

How It's Used

  • Replaying agent interactions to diagnose multi-agent system failures
  • Building reproducible datasets of agent behavior for reputation signals
  • Predicting outcomes of agent coordination strategies before production rollout
Works With
neo4jollamavuepython
Topics
aimulti-agentneo4jofflineollamaopen-sourcepredictionsimulationswarm-intelligencevue
Similar Tools
mirofishagent-playground
Keywords
multi-agentmulti-agent trustagent-to-agent evaluationneo4j