All posts

The agent needed a map

What a complex troubleshooting deployment taught us about procedural agents in production

In short

  1. Troubleshooting agents can’t afford ambiguity. With eight device types, 80 to 100 faults each, overlapping paths and live API checks, a conversation could sound natural and still break the procedure.
  2. Putting the whole tree in the prompt drifted, and one tool call per step was slow and rigid. Chunked context sat in between: the agent sees the part of the map it needs now.
  3. The result was a Guided Resolution Engine. Every fault is a graph; the agent moves freely inside its chunk, and the system re-anchors it when the caller goes backwards or changes the topic.
  4. Working from the partner’s offices meant same-day changes and no handoff: by launch, their team owned a system they had helped shape.
Published on Wonderful
Back to top