- Subject
- Windows & PowerShell
- Filed
- Source
- Suhasini Srinivasaragavan
- Loans
- 581
Pretxt Raises $5m to Supervise AI Agents While They Work
Pretxt’s $5m seed will support its runtime supervision platform, which aims to guide workplace AI agents when they depart from approved processes.

Pretxt has raised $5m in seed funding for a platform designed to monitor AI agents as they carry out workplace tasks and respond when their behaviour strays from approved processes. The company says its supervision layer evaluates agent actions in context, with the aim of addressing problems while work is underway.

That focus reflects a practical challenge for organisations deploying agents in customer conversations, transaction processing and other workflows: errors can involve skipped steps, invented exceptions or commitments the organisation did not authorise. Pretxt argues that reviewing activity after the fact leaves teams reacting too late. Its product is intended to intervene during an agent’s work instead.

Runtime supervision for workplace agents
Pretxt describes its platform as a supervision layer that assesses an agent’s behaviour against the context of its task. When an agent departs from an approved process, the system is intended to guide it back. The company also says the platform can identify areas where supervision could be strengthened over time.

The distinction matters for teams weighing how to extend automation into work with consequences for customers or business operations. A written rulebook may describe what an agent should do, but teams still need a way to observe whether it follows those instructions in a live workflow. Pretxt’s approach is to put that oversight into the runtime process, rather than rely only on reviews after a task is complete.

CEO Brian Soby says agents can do things they shouldn’t without a prompt injection attack. His point is that unintended behaviour need not begin with an obvious attack: agents may also make mistakes while handling ordinary work. Pretxt’s stated aim is to address those departures without blocking the work agents are meant to perform.

The company claims its platform provides supervision “60-times faster.” The announcement does not specify the comparison behind that figure, the test conditions or how the speed was measured. IT leaders assessing the claim will need those details to understand what it means for their own workflows.

Seed funding and early access
Pretxt announced the $5m seed raise on 17 November 2026. Deciens Capital, Howdy Partners and Service Provider Capital supplied the funding. The company is now available through an early access programme, though no further details about eligibility or availability were provided.

For organisations considering an early evaluation, the announcement establishes the product’s intended role but leaves operational questions open. It does not describe supported agent platforms, integrations, configuration requirements, pricing or how teams define the approved processes used for evaluation. Those details will determine how easily the supervision layer fits into existing workflows and what preparation an evaluation requires.

Teams will also want to understand what happens when the system detects a departure: whether it guides the agent automatically, alerts a person or uses different responses for different situations. Pretxt says the platform guides agents, but hasn’t described those controls in detail. The stated goal is to change an outcome while work is in progress, which makes the timing and form of intervention central to assessing the product.
Founders bring security backgrounds
Soby and CTO Aaron Costello co-founded Pretxt earlier in 2026. They lead the company from Dublin, Ireland, and Colorado, US. Both have careers in security, experience the company is bringing to the problem of supervising agents across workplace processes.
Soby previously co-founded California-based network security firm AppOmni, where he served as CTO until earlier in 2026. Costello spent several years as AppOmni’s chief of SaaS security research. He graduated from Technological University Dublin and has also worked at VMware’s Cork location and HackerOne.
Costello says supervising an agent shouldn’t require an organisation to become expert in every system the agent touches or in how the agent was built. That’s a useful test for the product: oversight has to fit the work and the organisation’s existing expertise, or it could add another layer for teams to manage.
Deciens Capital partner Vishal Rana described runtime supervision as a potential foundational layer for enterprise AI infrastructure, arguing that organisations need ways to influence outcomes before work is finished. That is an investor’s view of the market, rather than evidence of the product’s performance in customer deployments.
Pretxt’s early access programme gives prospective users a chance to examine the core proposition: whether contextual supervision can steer an agent during a task while allowing it to continue useful work. For IT teams, the next questions are concrete: which workflows the platform supports, how its interventions are controlled, and how its speed claim holds up under clearly described conditions.