Research
AI agents for high-stakes work
What we learned building Vextrum and Kairos.
An agent can prepare a convincing assessment while leaving out the fact that would change the decision. A reference beside an answer gives the reader somewhere to look, but still leaves them to establish whether the source supports the claim.
Drawing on Vextrum and Kairos, this paper sets out ten design requirements and five levels of trust, examines how evidence is checked, and explains where human judgement and approval remain essential.
Published
November 2026
Length
25 pages, PDF
Author
Sebastian Kaczmarczyk, Chief Product Officer
Inside
- The costly mistakes of AI agents read like correct answers
- We hold our agents to ten rules
- Trust in an agent comes in five levels, set by what the answer shows
- A trustworthy answer shows what it rests on
- Only questions the agent has never seen count
- Most of what makes an agent dependable lies outside the model
- Vextrum answers with its evidence attached
- Kairos shows every rule of a call in the call’s own words
- The person decides, and the system is built to keep it that way
- Agents are moving from answers to finished work
- Ten questions we ask of our own agents
- The hardest questions concern what the evidence can establish
Look inside
Cover
Page 8 · Five levels of trust
Page 16 · Evidence and judgement
Page 18 · The applicant decides
