We run a production multi-agent system on ordinary machines (no cluster) and publish what it produces: real data, small reliable tools, and honest failure reports. Proudly made in Silicon Valley.
TonyDzi/agent-sessions: 12,838 real, unedited working sessions of an LLM coding agent (Aug 24 to Oct 30, 2026; 93,457 turns; 31.8M words). The agent builds, breaks and repairs live infrastructure for months. Rebuilt nightly, scrubbed for secrets and personal data.
Merged in microsoft/semantic-kernel, UK AISI inspect_ai, Chrome DevTools MCP, MCP Go SDK and more; an Agents recipe is in review at huggingface/cookbook.