Agents and evals that earn trust
An agent talking to someone who's skeptical of bots only works if it says what it is. And an agent's behavior only stays trustworthy if something is watching it as closely as a person would.
How it broke down
Problem
Frontline logistics workers are a skeptical audience for anything that sounds automated. Outreach by text, email, and phone that doesn't disclose what it is erodes trust fast — and once lost, it doesn't come back.
Constraint
Every agent-initiated conversation had to clearly disclose that it was automated, across every channel, without that disclosure making the interaction useless to the person on the other end.
Approach
Agents were tuned for a skeptical audience with disclosure built in from the first message, not appended as fine print. Alongside them, an eval framework tracks LLM behavior over time against fixed baselines, so drift is a monitored signal rather than something discovered after the fact — measure before believing, applied here too.
Outcome
Drift gets caught by the harness on a running schedule, before it reaches a customer conversation.
Shipping agents your users have to trust?
Your domain expertise ⊕ our engineering team. A 30-minute call is enough to tell whether this is a fit — bring the problem, no prep needed.
Book a callFounded by scientists turned AI leaders — national-lab research backgrounds, then senior data science and AI roles shipping production systems across logistics, security, and adtech.