Open-source agent harness runs 63-hour autonomous Riemann hypothesis attempt on a single RTX 3090
GuiltyBookkeeper4849 · reddit · 2026-09-14
A Reddit developer open-sourced Artificium, a general agent harness for long-running autonomous work, and stress-tested it with an extreme experiment:
- Setup: exactly two messages (start, then request final report) drove the agent to autonomously work on the Riemann hypothesis for 63 hours, processing 50M+ tokens on a 4-bit quantized Qwen3 27B on one RTX 3090.
- What it did: explored approaches, wrote and debugged programs, built long-term memories to continue across sessions—without hallucinating a fake solution or losing focus.
- Harness features: self-managed working/long-term memory, tool building, self-revision of purpose and harness, non-blocking multi-user interaction, agent-to-agent communication, sensor inputs.
- Open release: harness on GitHub; the experiment's memories, programs, life-loop logs and metrics published as a Hugging Face dataset. The author, inspired by OpenAI's Navier–Stokes autonomous research, plans a livestreamed open-model agent swarm attacking a Millennium Prize Problem.
More from coding & agent
- Seekr MCP Server Adds Google/Wikipedia Search and Webpage Extraction — modelcontextprotocol · 2026-09-14
- Wikipedia Trends MCP Launches: Page View Growth Data for Any Article — modelcontextprotocol · 2026-09-14
- Building an Agile ERP Chatbot: Developer Seeks Advice on Intent Routing — Neither-Owl-1772 · 2026-09-14
- All seven Hermes Agent release explainers collected on one page — Teknium · 2026-09-14
- Open-source macOS client Locus 3.0: one GUI for ChatGPT, Claude and Kimi plans, plus planning-executing agent duo — drankthedew · 2026-09-14
- Prompt Library: A Folder-Based Prompt Organizer Node for ComfyUI — MediocreRegister9642 · 2026-09-14