Workflow Mutation
Generating workflow variants — different prompts, models, retrieval policies, node shapes, tool budgets — for the eval harness to score before any change reaches production.
A workflow change shipped without measurement is an opinion enforced by deploy. Mutation makes the change a candidate the harness can score.
What it solves
Turns workflow optimization into a measurable, reviewable process. Surfaces dominant candidates without trial-and-error on production traffic.
How we build it
Mutation operators per surface: prompt rewrites, model swaps, retrieval-depth changes, node restructures, tool-budget tweaks. The harness runs the cartesian product (or a guided search) against the eval set and reports Pareto-dominant candidates.
- Per-surface mutation operators
- Guided search across surfaces
- Pareto-dominant candidates reported
- Human approval before promotion
What changes when it is in place
Workflow improvement becomes data-driven. Changes that survive the harness are the ones that ship; the rest stay in the variant table.