Case Study - An agentic assistant for the field service trades

Overview

Field service professionals — plumbers, electricians, HVAC techs — spend their day with their hands full and their attention on the job, not on software. The goal of this system was to let them run the business side of their work the way they'd talk to an office manager: by voice or chat, in plain language.

Under the hood that meant a real agentic system, not a single prompt. A customer agent, an estimate agent, a scheduling agent, an invoice agent and several others each own a slice of the domain. A coordinating graph routes the conversation to the right agent, maintains durable state across turns, and keeps shared context — the customer, the job, the day's schedule — consistent throughout.

What we did

We treated agents as the owners of a conversation and reusable subgraphs as the domain logic they compose — so the customer-collection flow, for example, is written once and reused by every agent that needs it. Long-term memory, short-term checkpointing, and working state each use the right primitive rather than a custom store bolted on.

A core safety decision: agents can create, read, and update, but destructive actions stay in the UI. A voice assistant that can delete an invoice on a mishearing is a liability, so the architecture draws that boundary explicitly.

Voice that works on a job site

The voice layer is transparent to the agents: speech is transcribed, cleaned of filler, and handed to the same backend that serves text. Responses stream back sentence by sentence into text-to-speech so the assistant starts talking immediately instead of waiting for a full answer, and the user can interrupt at any time — the way a real conversation works.