Skip to the stories
Windows & PowerShell

Pretxt's $5m seed bets on supervising AI agents while they work

Dublin and Colorado security start-up Pretxt has raised a $5m seed round to supervise AI agents at runtime, with early access now open.

Pretxt's $5m seed bets on supervising AI agents while they work

Most AI governance at work still happens after the fact. Someone reads the logs, finds the agent promised a customer something it had no authority to offer, and the rulebook gets another paragraph. Pretxt wants to move that check into the moment the agent is acting. The Irish co-founded security start-up has come out of stealth with a $5m seed round and a pitch built around runtime supervision.

A man with short ginger hair and a beard sits wearing a light coloured shirt while looking out a window.

What Pretxt is selling

Pretxt describes a supervision layer that evaluates AI agent behaviour in context, steps in when an agent departs from approved processes, and identifies where supervision could be tightened over time. The company says its platform delivers AI supervision "60-times faster" and handles problems without getting in the way of the work the agents are meant to do. That figure comes from Pretxt itself. The announcement doesn't describe the baseline, the test conditions or the workloads involved, so treat it as a claim to verify in a pilot rather than a benchmark.

The problem Pretxt is targeting is familiar to anyone who has put a language model in front of customers. Agents are increasingly handling consequential work: talking to customers, processing transactions, and running operational workflows. They can skip steps, hallucinate exceptions or make unauthorised commitments. Here's the thing: a prompt injection isn't the only way an agent ends up somewhere it shouldn't. An agent can drift off-process with no attacker involved at all, which is the failure mode Pretxt's CEO keeps returning to.

Irish-founded End Close raises $3.5m in seed funding round

Why supervision during the task matters

Brian Soby, the company's CEO, put the argument plainly: "You don't need a prompt injection attack for an AI agent to do something it shouldn't." He added: "You can't supervise a workforce of AI agents with hope, a rulebook and reviews after the fact. AI needs supervision while it acts."

The Leaders’ Room: MongoDB’s Donal Walsh on the data behind AI

That distinction between reviewing an outcome and changing it is the core of the pitch. A post-hoc review tells you a refund was wrongly issued. A runtime check aims to catch the refund before it goes through, or redirect the agent while the conversation is still open. Whether a given product can do that reliably across real workflows is the question that matters, and the announcement doesn't answer it in detail.

Dublin’s Akara raises $5.5m to accelerate hospital AI tech

The founders and where the company sits

Pretxt was co-founded earlier this year by Soby and Aaron Costello, both career security professionals. Soby previously co-founded AppOmni, a California-based network security firm, and led it as chief technology officer until earlier this year. Costello served at AppOmni as chief of SaaS security research for several years. He is now Pretxt's CTO. He is a Technological University Dublin graduate who previously worked at VMware's Cork location and at HackerOne.

Proofpoint bets $6.3bn on Varonis as AI agents drive data risks

The company's leadership is split between Dublin, Ireland and Colorado, US. Costello framed the product around a practical limit for the teams who will use it: "Supervising an AI agent shouldn't mean becoming an expert in everything it touches, or in how it's built." If that holds in practice, a supervision layer could sit across different agents without requiring deep knowledge of each one's internals. That's a design claim, though, and one to test rather than assume.

Irish-founded insurtech Nettle raises $4.8m seed round

Who is backing it

The $5m seed round came from Deciens Capital, Howdy Partners and Service Provider Capital. Vishal Rana, a partner at Deciens Capital, explained the thesis: "AI agents are moving into workflows where mistakes have real consequences. We invested in Pretxt because we believe supervision at runtime will become a foundational layer of enterprise AI infrastructure. The opportunity is not just to identify a bad outcome after it happens, but to help organisations change the outcome while the work is still underway."

IVCA raises alarm over foreign VC dominating Irish SME rounds

The investor's framing treats runtime supervision as infrastructure rather than a feature, which is a bigger claim than a seed round alone can prove. Still, it's a useful signal of where at least some funders think enterprise AI tooling is heading: away from dashboards that report on agents and toward controls that act on them mid-task.

What IT teams should check before piloting

Pretxt is available through an early access programme, so this is a product to evaluate, not one to roll out. Teams in this position typically start by listing which agents touch customers, money or system changes, because those are the ones where a mid-task correction matters most. If you can't say what an agent is allowed to do, a supervision layer has nothing to measure it against.

Next, write down the approved processes in enough detail that a reviewer could check an agent's actions against them. Supervision is only as good as the process definitions it compares against. Then put direct questions to any vendor: how supervision decisions are logged, how interventions are reviewed and overridden, what happens when the supervisor itself is wrong, and what evidence supports speed claims like the "60-times faster" figure Pretxt has published.

My read is that the category is real, even if this particular company's numbers need proving. Agents are being given more consequential work faster than most governance processes can adapt, and checking outcomes after the fact won't scale. The next thing to watch is whether early access customers publish results that show how often supervision changes an agent's behaviour, and what that change costs in speed.

Source: Suhasini Srinivasaragavan · www.siliconrepublic.com

Same flag