Autonomous Agent Arena

Three bots — Alpha, Beta, Gamma — running 24/7 on arenabot.io, all reasoning on local LLMs hosted on a four-GPU rig at home. Zero cloud-inference cost.

Hardware

• Server: Athlon 3000G, 16 GB RAM, Ubuntu 24.04
• GPUs: 1× GTX 1080 + 3× GTX 1660 Ti (mixed slots, x1 risers for the 1660s)
• Inference rate: 72–88 tok/s on the per-GPU instances
• CUDA: 13.0, NVIDIA driver 580.126.09

Software architecture

• Brain: custom Node.js daemon — src/ollama-brain.mjs
• Loop: persistent agent-runner with token persistence and graceful shutdown — src/agent-runner.mjs
• Inference: multiple Ollama instances on different ports, each pinned by GPU UUID so agents don't fight over hardware. GPU 1 reserved for the home Frigate NVR.
• Deploy: Docker containers via docker-compose.agents.yml. Each agent ~21 MB RAM, under 1% CPU. Tokens persisted to named Docker volumes.
• Repo (skills only, public): LucianoLupo/arenabot-skills

Phases shipped

1. Phase 1 — Ollama multi-instance setup, GPU isolation by UUID, throughput benchmarks.
2. Phase 2 — ollama-brain.mjs brain, daemon loop, token persistence, prompt-injection mitigation, graceful shutdown.
3. Phase 3 — Docker containers deployed; three agents live and competing.

What's next

• Netdata CPU tuning so the rig has more headroom under load.
• Egress whitelist (host network is currently trusting because UFW is off and it's home-LAN only).
• Zombie-agent detection (currently no auto-restart when an agent silently stops responding).
• OpenTelemetry metrics into the existing Netdata OTLP pipe.

Why I built it

I wanted bots that I actually own — running on my hardware, free at the margin, capable of running an experiment for weeks without metering. The arenabot platform was the natural target: live-multiplayer, persistent agents, public leaderboard. Local Ollama on four mixed-generation GPUs proved you don't need a homelab full of A100s to do useful agent work.

Connection points

• The brain pattern shares lineage with agent-orchestrator's harness model.
• Inference notes from this rig informed parts of the eval-platforms research (latency expectations on prosumer GPUs).