Optimization loop

This is Donkit's self-optimizing core. Once you can measure quality, the platform can propose better configurations of your agent — and prove they're genuinely better before you adopt them. Optimization is plan-gated, and only one optimization runs per agent at a time.

How it works

You start an optimization from a completed dataset evaluation. Donkit generates candidate versions of your agent, evaluates each against the same yardstick as your baseline, and keeps only candidates that genuinely improve on it. You set an improvement threshold so trivial or noisy gains don't count.

The method

Knowledge / RAG optimization tunes how your agent retrieves from its knowledge — how many passages it pulls per question and how it phrases the lookup — and, in a fuller mode, how your documents are split when they're indexed. Each combination is scored against your dataset, and only a combination that beats your baseline is offered.

The optimization tree

Every optimization is visualized as a tree of steps: the baseline at the root, each candidate branching from it, annotated with its scores. You can watch it build live, expand any candidate to see what changed and how it scored, and stop a run at any time.

Promote or reject

Candidates are compared to your baseline using your improvement threshold; only those that clear it count as real improvements. You review the winning candidate — its score delta and what changed — and decide whether to adopt it as your agent's new version.

The platform proposes and proves — you approve

Donkit reliably finds and proves better configurations and surfaces them for adoption. It presents winners for your review rather than silently replacing your agent, so you stay in control of what ships.