Eigen RadarAI
Analysis

Pretxt opens early access to supervision for working AI agents

US and Ireland-based Pretxt has introduced AI Supervision for checking automated agents while they carry out business tasks. Teams describe their rules in ordinary language and approve changes to supervision plans. The early-access product aims to redirect agents that stray beyond their authority, including customer-service systems making unauthorized commitments.

Artificial Intelligence··Night
Two enterprise reviewers in a Dublin software studio examine a colour-coded supervision workflow chart at a high workbench, illustrating human review and redirection of AI agents.

Supervision checks agents during work

Pretxt, an enterprise software company based in the US and Ireland, opened early access to AI Supervision on November 17. It targets AI agents: programs that use tools and perform tasks on an organization’s behalf. Pretxt says its supervision layer checks behavior while work is underway and redirects an agent when it departs from the organization’s instructions or authority boundaries.[1], [2]

Teams define boundaries in ordinary language

Enterprise teams describe expected behavior and limits in natural language. They can test a proposed supervision plan against their own examples before approving it. Changes suggested from observed behavior also require human review. Pretxt describes an engine that does not ask another large language model to judge every action. Its co-founders, Brian Soby and Aaron Costello, previously worked at AppOmni.[1]

Refunds and verification are among intended uses

The company’s examples include stopping unauthorized refunds, keeping marketing offers within approved terms and ensuring borrower verification in financial workflows. Healthcare applications concern scheduling and information within clinical boundaries; insurance exceptions and high-value decisions can go to human approvers. These are proposed uses of the early-access system. Organizations are invited to bring their own workflows and operating requirements for evaluation.[1]

References

  1. News sourcePretxtPretxt opens early access to runtime agent supervision↩1↩2↩3
  2. News sourceSecurityWeekPretxt introduces runtime supervision for AI agents↩