Reasoning agents

Runs the same task num_samples times, independently and in parallel, then hands every answer to an aggregator agent that compares them and returns the answer most of them agree on. Disagreement between samples is a signal that the question is hard or ambiguous; agreement makes the final answer far more reliable than a single run. This is the default reasoning agent type.

Answers, critiques itself and improves. Each loop it drafts or refines a response, scores it out of 10, writes a reflection on what to fix, and feeds that reflection into the next revision. Reflections are kept in short and long-term memory, and the best-scoring response across all loops is returned, so a weaker late revision never replaces a stronger early one.

A judge for outputs you already have. Pass the output to evaluate as the task; the judge identifies the domain, applies criteria that fit it, and returns a reasoned assessment of strengths, weaknesses, factual problems and concrete improvements. With max_loops above 1, each pass sees its earlier verdicts and sharpens them.