Agent societies that vote, remember, and specialize—unsupervised

What it does

Ava is a headless Python framework for running hundreds to thousands of autonomous agents in a shared environment. Each agent owns a tiered memory system, a belief model, and a social graph, stepping through a perceive-deliberate-act loop every tick. The goal is to watch civilization-level properties—specialized roles, shared norms, emergent laws—arise from local interactions without being scripted in advance.

The interesting bit

Rather than trusting a demo video to prove emergence, Ava ships an analysis layer that runs Mann-Whitney U change-point detection over rolling population metrics, turning “it looks like they formed a government” into a statistically backed claim with a p-value.

Key highlights

• Memory is split into episodic, semantic, and procedural stores with distinct retrieval and decay rules; successful skills reinforce independently of any single episode.
• Ticks are scheduled via asyncio.TaskGroup under a bounded semaphore, so populations scale without unbounded coroutines or one slow agent freezing the world.
• Agents can propose and vote on laws through a GovernanceSystem whose quorum scales with population, and a three-phase BFT consensus protocol handles adversarial participants.
• A reference dashboard inspects live agent state, memory, and relationships, built on SimulationTracer so the core engine stays frontend-agnostic.
• Benchmarks show throughput holding around 20–25k agent-steps per second for populations up to 5,000.

Caveats

• The README truncates its own performance table mid-line, leaving the 5,000-agent p95 and exact throughput incomplete.
• The consensus module is named RaftLikeConsensus but implements a PBFT-style three-phase commit, which may mislead readers searching for actual Raft logic.

Verdict

Worth a look if you are building large-scale social simulations or multi-agent reinforcement environments and need statistical rigor around emergence. Skip it if you just want a quick LLM chatbot wrapper—this is a systems framework, not a conversational agent.