RamLab.dev
← Lab

Local AI research

Evolution Lab

An ongoing experiment in what local AI agents can learn and what their run evidence can actually establish.

Evolution Lab runs local AI agents across repeated sessions. The agents use language models served through LM Studio, maintain state between cycles, and can record hypotheses, actions, and observations. The project asks how much follow-through this setup produces and how to measure it.

A four-hour run

On November 21, 2026, three agents ran with qwen/qwen3.8-27b through LM Studio. This was an ordinary run without a seeded benchmark task.

  • Agent 1 completed 11 cycles, Agent 2 completed 11, and Bob completed 10.
  • The report recorded 31 API errors across the three agents.
  • The synthetic preflight produced 3 of 3 valid outputs. Longer calls still encountered errors during the run.
  • The run recorded zero independently verified capability improvements.

What I track

An agent cycle is a unit of runtime activity. A task closure needs a machine-checkable assertion backed by evidence. A capability improvement needs a comparison against a baseline on relevant, independent cases. Those measures answer different questions, so they appear separately in the run report.

What I’m investigating

I’m looking at how reliably an agent can turn an investigation into a testable task, carry it through to a recorded result, and recover when the model produces an incomplete response.