Pretxt launches from stealth with $5 million in seed funding. Its tool supervises AI agents at runtime, stepping in before mistakes land, and targets regulated industries first.

Security startup Pretxt has launched from stealth with a $5 million seed round, selling a layer that watches AI agents while they work and steps in before their mistakes land

1

. The round was announced on 17 November

1

.

The timing follows a broader shift the outlets describe: companies are trusting AI agents with consequential work such as refunds, database changes and customer transactions, which is exactly the kind of work where an error costs real money

1

2

.

Why the founders built it

The founders argue that the two obvious fixes can both fail: blocking an agent outright slows its work down

1

2

. Flagging mistakes after the fact catches them too late to prevent damage

1

2

. Agents can skip steps, invent exceptions or make unauthorized commitments on their own, without any prompt-injection attack

1

2

. "You can't supervise a workforce of AI agents with hope, a rulebook and reviews after the fact," said CEO Brian Soby. "AI needs supervision while it acts"

2

.

The pair got a close look at the problem at their previous company, AppOmni, where they uncovered major vulnerabilities in AI agents built by Salesforce and ServiceNow

1

2

. Soby left AppOmni's CTO role earlier in 2026 and previously held senior security roles at Salesforce and Mitre; CTO Aaron Costello previously worked at VMware and HackerOne

1

2

.

Source: SiliconANGLE

Source: SiliconANGLE

What the product does

Pretxt's tool, which the company calls AI Supervision, evaluates what an agent is doing against the company's own rules and policies and intervenes in real time to steer it back on course

1

2

. Teams write those policies in plain language, and when an agent hits a situation nobody anticipated, the tool halts it and escalates to a human

1

.

The company says its engine is up to 60 times faster than tools that use a second AI model as a judge, because it is near-deterministic rather than another language model weighing each action

1

2

. On one benchmark, Zapier's AutomationBench, violations reportedly fell from about one in five agent actions to zero, according to the company

1

. The evidence so far is thin outside the company's own walls: the benchmark result is company-reported, from a single test, and no independent validation is cited

1

.

The product is in early access, aimed first at regulated, high-stakes industries: verifying borrowers and routing complex claims in financial services and insurance, keeping agents within strict boundaries in healthcare, and blocking unapproved discounts or refunds in customer support

1

2

.

The bet behind the round

Deciens Capital led the seed round, with Howdy Partners, Service Provider Capital and NSV Wolf Capital participating, according to SiliconANGLE

1

. Silicon Republic's account lists the same investors minus NSV Wolf and names no lead, so the full lineup rests on one report

2

. One analyst, Chris Steffen of Enterprise Management Associates, says it is effectively a new category of tool, because it checks whether agents are doing their assigned job correctly rather than only checking security or compliance

1

. "Most tools can only report what went wrong after the fact, when it's too late," he said. "Pretxt intervenes while the outcome can still change to get agents back on track"

1

.

"We invested in Pretxt because we believe supervision at runtime will become a foundational layer of enterprise AI infrastructure," said Vishal Rana, a partner at Deciens Capital

2

.

© 2026 TheOutpost.AI All rights reserved