The situation

The product team decides that a user can now belong to several workspaces. The spec is two paragraphs. The test repository is 1,400 cases, written over three years by people who never used the phrase “multi-workspace”.

What the agent does

  1. Calls get_project_rules to load the project’s vocabulary and constraints.
  2. Calls smart_search_cases with the behaviour, not the feature name: “journeys affected when one user can join several workspaces”.
  3. Receives a ranked family of approved cases: invitation acceptance, ownership transfer, workspace-scoped deep links—each marked strong or related.
  4. Calls find_related_cases on the strongest hit to widen the net through the neighbour graph.
  5. Reads the full cases—steps and expected results, not just titles—before proposing a single change.

What the human reviews

A changeset with three columns: cases to update, cases to add, and intent that should remain untouched. The last column is the one generation-first tools never produce.

Why it works

Each case carries two embeddings: its own text and an agent-written semantic profile that restates the behaviour in user language, other phrasings, and technical terms. A query in defect language, product language, or engineering language lands on the same approved intent.

Try the instrument below with the three example queries.