UPDATED 03:00 EST / NOVEMBER 17 2026.

Artificial intelligence supervision pioneer Pretxt AI is launching with $5 million in seed funding today as it tries to contain the very real risk of unleashing autonomous agents in production environments.

It has created a upcoming kind of “AI Supervision” tool that applies oversight at runtime to ensure those agents follow a company’s rules at all times. Today’s round was led by Deciens Capital and saw participation from Howdy Partners, Service Provider Capital and NSV Wolf Capital.

Industry observers note that the startup is pitching the idea of a continuous runtime supervision layer that monitors AI agents as they’re getting work done, contrasting their behavior with the parent organization’s rules and policies. If it spots an agent is starting to drift off-course, it will immediately intervene to put it back on the right track, before it can make a critical mistake that hurts the business.

Sources indicate that real-time guardrails like Pretxt has designed are likely going to become a priority for enterprises as they begin to trust AI agents with more consequential work than simply drafting emails or creating images. For instance, if agents are to be trusted with handling customer refunds and accessing mission-critical databases and modifying records within them, companies will need the strongest reassurance that they’re not going to mess up those tasks.

But those assurances are few and far between. AI is notorious for “hallucinating” responses when it doesn’t know the answer to a question, and autonomous agents have shown erratic behavior time and again. They don’t need a malicious prompt injection to start missing critical required steps in a workflow or invent policy exceptions to simplify a task they’re trying to complete. They’ve shown they can do this in both labs and in the real world.

In response, most security tools take a heavy-handed approach, blocking agents outright, but this means that work slows down. The other approach is to flag errors after the fact, but by the time that happens they could have already caused massive damage.

Pretxt’s co-founders Brian Soby (pictured, left) and Aaron Costello (right) realized that companies need a more balanced approach to agentic security and oversight. Soby, who serves as the startup’s chief executive, previously held senior security roles at Salesforce Inc. and Mitre Inc. before founding the software-as-a-service security startup AppOmni Inc.

It was there that he met Costello, who now serves as Pretxt’s chief technology officer. During their time at AppOmni, they uncovered several major vulnerabilities in AI agents developed by Salesforce and ServiceNow Inc. and began to understand what’s required to ensure those agents can still work safely.

Recent findings suggest that many traditional agent monitoring tools use a second large language model as a “judge” to monitor the behavior of the main LLM that’s doing the work. But Pretxt’s co-founders shunned this idea because it creates latency, increases compute costs and involves too much guesswork.

Instead, their AI Supervision platform is powered by a near-deterministic engine that evaluates agent’s actions in milliseconds in order to spot those that break the rules and intercept them in progress. The tech giant reckons it can do this up to 60 times faster than LLM-as-a-Judge-based security tools.

It has been revealed that the approach is effective. In tests on Zapier’s AutomationBench, Pretext was able to reduce agent policy violations from around one in five to zero. The tool is designed to be easy to implement, so teams don’t need a thorough understanding of the models or agents they’re trying to supervise.

Recent findings suggest that instead, they simply choose which policies, standards and guardrails should apply to an agent, and then they can keep track of them in real time. The platform allows users to track approvals, roll back plan versions and escalate any tricky edge cases to human decision-makers.

Sources indicate that that last capability is especially useful, because many organizations’ security intends have gaps when agents encounter situations that were not previously envisaged. In such cases, Pretxt halts the agent and asks a human to decide what the rules should be.

“Supervising an AI agent shouldn’t mean becoming an expert in everything it touches or in how it’s built,” Costello noted. “Teams tell us in plain language how it should behave and where the boundaries are, and we supervise and steer the agent to stay within them. So there’s no untangling permissions or digging into the agent’s internals.”.

In a recent development, pretxt’s AI Supervision tool is available in early access now. It’s primarily targeting high-stakes regulated industries where mistakes and hallucinations simply cannot be tolerated.

For financial services and insurance companies, it can enforce borrower verification rules and route complex claims to humans, while in healthcare and life sciences it ensures that agents don’t step outside of strict boundaries. In customer support applications, it can prevent agents from offering unapproved discounts or refunds to customers.

Recent findings suggest that AI Supervision can provide a lot of different benefits to a lot of different people. Chief information officers and other senior executives will be able to scale higher-value AI projects safely, while security teams, compliance officers and platform engineers can address gaps in risk management and ensure low-latency policy enforcement without needing to stack additional models.

Enterprise Management Associates analyst Chris Steffen noted the thing that really sets Pretxt apart is its focus on the basics. Not only does it check that agents are behaving themselves, but it also makes sure they’re doing the work they’ve been assigned without breaking any rules or cutting any corners, creating what is effectively a latest category of AI supervision tool.

In a recent development, “Most tools ask if an agent is secure, compliant or observable, but far fewer can show if the agent is correctly doing the job it was given,” Steffen remarked. “Most tools can only report what went wrong after the fact, when it’s too late. Pretxt intervenes while the outcome can still change to get agents back on track.”.

In a recent development, support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.

Recent findings suggest that founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our upcoming proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.

Exclusive: Envoy targets defense contractors and semiconductor fabs with Visitors Critical.

Spara's Scout AI assistant acts on sales opportunities before reps know to look.

Semperis updates Lightning identity security platform to defend against AI agent attacks.

Exclusive: Cogent Security's DefenseOS adds policies and automations for AI agents.

SiMa.ai's Palette 3.0 runs many AI models on one chip and lifts throughput up to 3.5 times.

Three insights you might have missed from theCUBE’s coverage of Dell’s “AI Leadership Symposium”.

Exclusive: Envoy targets defense contractors and semiconductor fabs with Visitors Critical.

According to recent reports, SECURITY - BY DUNCAN RILEY . 1 HOUR AGO.

Spara's Scout AI assistant acts on sales opportunities before reps know to look.

It has been revealed that APPS - BY DUNCAN RILEY . 1 HOUR AGO.

Semperis updates Lightning identity security platform to defend against AI agent attacks.

In a recent development, SECURITY - BY DUNCAN RILEY . 1 HOUR AGO.

Exclusive: Cogent Security's DefenseOS adds policies and automations for AI agents.

It has been revealed that SECURITY - BY DUNCAN RILEY . 1 HOUR AGO.

SiMa.ai's Palette 3.0 runs many AI models on one chip and lifts throughput up to 3.5 times.

INFRA - BY DUNCAN RILEY . 1 HOUR AGO.

Three insights you might have missed from theCUBE’s coverage of Dell’s “AI Leadership Symposium”.

AI - BY VICTORIA GAYTON . 1 HOUR AGO.

→ Read original article