I want an Agent to organize addresses stored in Busabase into province, city, and district fields. Some addresses contain abbreviations or incomplete information. How should I describe the task so that clear cases can be processed while uncertain ones are left for review? I would like the results to explain what could not be determined rather than fill the gaps with guesses.
How should I instruct an Agent to handle uncertain data?

Start by making uncertainty a valid output, not a failure the agent must hide. In a Busabase Addresses Base, keep the original address unchanged. Add province, city and district as separate optional fields, plus a review status such as confirmed / needs review, a reason, and the source or rule used. The agent should fill a component only when the source supports it; an ambiguous abbreviation must not be expanded by guesswork. A missing district stays empty with a reason, rather than being inferred from the city alone.
I would give the agent this instruction: “Read a small batch of address records. Preserve each original address. Split province, city and district only where the text or an approved reference resolves them unambiguously. For every uncertain part, leave that field blank, mark needs review and explain the competing interpretations or missing detail. Return record links and proposed changes before writing. Do not overwrite confirmed values without showing why.” Give it a review-only credential so the team can compare the proposals before merging.
Test ten varied addresses first, including abbreviations and missing districts. Review both false matches and records flagged for help, then save the agreed mapping rules as a workspace Skill for the next batch. Do you have an authoritative region list or existing customer records the agent may use to resolve abbreviations?
Your reply
Sign in to join the discussion. Reading is always open; you only need an account to post.
Sign in