Skip to content

Product case · 1 / 13

When an agent should stop

Product note. Organizations, dates, private data and implementation details are omitted.

Abstract illustration: a funnel of lines that stops before its opening, with one line breaking away

The problem and the decisions

I explore how a chat assistant can recognise its limits and hand over to a person.

The decision: define what it can resolve, what it should ask, and when it should stop.

What it was built with

  • Claude API
  • FastAPI
  • Supabase
  • Railway
  • ManyChat/Meta

What I did

  • Agents
  • Human handoff
  • Testing the answers
  • Testing

What I learned

A confident answer is not always a helpful one.

More context · password required →