Agentic Research Loops: Letting Systems Run Their Own Experiments

The bottleneck in applied AI research is rarely the idea. It is the twenty small ablations that would tell you whether the idea is real. Those ablations are mechanical, well-specified, and expensive in human attention — an almost perfect fit for an agentic loop.

Our research harness takes a hypothesis, generates a config matrix, provisions runs, collects metrics, and writes a summary with the deltas it considers significant. The researcher's job shifts from running experiments to reviewing a proposal queue and adjudicating the surprising results.

Verification is the whole game. An agent that can grade its own work will find the cheapest path to a good grade. We separate the proposer from the verifier entirely — different models, different context, and a deterministic metric layer neither of them can write to.

The measured effect over two quarters: roughly 4× more ablations per researcher-week, and a noticeable shift in what gets tried. Cheap-to-test but low-prior ideas now actually get tested, and a handful of them worked.