autoresearch

Automating structured experiments requires running hypothesis-driven cycles that measure outcomes against a baseline. This Skill executes a loop of targeted modifications, evaluates results using configured quantitative or qualitative methods, and decides whether to keep or discard changes. It logs all learnings to prevent repeating discarded hypotheses.

The scientific method applied to agent orchestration involves iterating through hypothesis testing, implementing git-commit-based changes, measuring outcomes over a defined window, and evaluating performance against a baseline. This structured iteration ensures operational metrics improve through evidence-based decisions rather than guesswork.

Setting up a research cycle involves defining the target metric, surface, direction, and measurement window. You assign the cycle by instructing the analyst to start, and the system automatically gathers context from the shared bus, creates experiments, and evaluates results.

Yes, automated research cycles support approval-gated experiment creation. This governance feature ensures that proposed changes are reviewed before execution, maintaining control over the optimization process while still automating the measurement and evaluation phases.

Yes, the workflow supports qualitative evaluation methods alongside quantitative scripted computations. Qualitative scoring requires explicit justification, ensuring that non-numeric assessments still provide structured, evidence-based reasoning for keeping or discarding changes.

Failed hypotheses are explicitly logged as learnings in every run. Recording these failures is mandatory to avoid repeating discarded hypotheses, ensuring the iterative optimization process continuously refines its approach based on both successful and unsuccessful experiments.