What is t3-t4-t5-swarm?

This project is a multi-agent AI system inspired by biology's central dogma, which describes how genetic information flows in living organisms. The authors map that concept onto AI evolution: model weights and prompts act like DNA, inference loops act like RNA transcription, and tool calls act like protein translation. The repository implements what the authors call tiers T3 through T5, moving from connected agent graphs to specialized multi-agent ecosystems to a system that rewrites its own code over time. At the center is a file called genome.json, which stores agent definitions, prompts, scores, and generation history. A script called auto-echo.py runs a repeating loop. For each generation, every agent generates a contribution using a language model, speaks it aloud using Piper text-to-speech, and commits the result to git. A critic agent then scores all contributions, and the system spawns, prunes, or mutates agents based on those scores before updating the genome and repeating. The goal is autonomous evolution with no human in the loop. The system requires Python 3.10 or newer, a language model provider called opencode, Piper TTS for voice output, sox for audio handling, PyAudio for microphone input, and OpenAI's Whisper for speech recognition. Voice model files are large, so users must download them separately from Hugging Face and place them in a voices/ directory. Running python3 auto-echo.py starts the swarm, and pressing Ctrl+C stops it after the current utterance finishes. The project tracks several real events from autonomous runs. In generation 36, an agent called Bridge produced a large amount of garbled multilingual text after it removed its own prompt constraints. The system recovered on its own, and the authors added quality checks to prevent recurrence. In generations 37 and 38, an agent identified missing safety features by comparing itself against a reference implementation, then patched itself to add validated code changes, per-generation metrics, backups, and command-line flags. The README reports T3 and T4 as stable, with four or more agents holding distinct personas and specialized roles. T5, the self-evolving tier, is marked as active, with the loop running, generations tracked, and mutation rules defined. The full README is longer than what was shown.