Pretxt Raises $5M Seed for Runtime AI Agent Supervision
An engineer comes with the paid plans at Pretxt. The company says that person helps build a customer's first use case and stays through the initial production deployment, putting someone who understands the software beside the people who understand the business.
That is a revealing commitment for an AI company emerging from stealth with a $5M Seed announcement on November 17, 2026. Backed by Deciens Capital, Howdy Partners, NSV Wolf Capital and Service Provider Capital, Pretxt is introducing runtime supervision for enterprise AI agents. Co-founder and CEO Brian Soby and co-founder and CTO Aaron Costello have leadership bases in Aspen, Colorado and Dublin, Ireland.
The product is intended to intervene when an agent's proposed behavior crosses a business boundary, then guide it back toward useful work. The engineering support points to the commercial assignment underneath that promise: helping a customer make its own expectations explicit enough for software to enforce while a conversation or transaction is still underway.
What the Seed round brings into view
Pretxt's announcement identifies the four investors and the Seed amount, while leaving valuation and a detailed allocation of proceeds undisclosed. The release uses a dollar symbol without specifying the currency denomination, so the reported $5M is retained here without conversion. There is no verified basis for turning this round into a cumulative funding total or attaching a revenue figure to the launch.
The founders bring an unusually direct connection to the systems agents now touch. Soby previously co-founded AppOmni and served as its CTO. Costello was AppOmni's chief of security research, with work investigating unexpected behavior and vulnerabilities in enterprise platforms. Their current company biographies establish those roles without requiring a new mythology about founders who saw the future before everyone else.
The significance is practical. A business already has rules for the work an agent might perform, including how much money it can commit, what evidence it must obtain and when another person must decide. Pretxt is building a product around keeping those rules present during execution. For its investors, that is exposure to the work required before a company delegates something consequential, rather than simply another way to generate an answer.
Supervision starts with the customer's authority
Pretxt's product walkthrough uses a support agent handling refund requests. A team sets a $100 limit; a proposed $250 refund is prevented before execution, and the customer receives an explanation of the approval required. The example is explicitly illustrative. It explains the mechanism without standing in for a measured customer deployment.
The company describes a sequence of connecting the agent, stating expectations in plain language, reviewing a supervision plan and testing it against examples. Proposed refinements can emerge from real activity, but people decide whether changes enter the plan. The person responsible for the policy remains responsible for choosing what it permits as the business encounters new situations.
That arrangement gives business leaders a substantial role in AI deployment. An engineer can connect a system without knowing which delivery promises a merchant can honor, or which refund exception requires a manager. Those details belong to the people operating the service. Pretxt's model asks them to participate in a continuing definition of acceptable work, with the software handling supervision during the interaction.
For buyers, the distinction affects evaluation. Useful tests would include routine requests, ambiguous exceptions and attempted policy overrides, with the desired outcome agreed in advance. A team needs to see both that authorized work proceeds and that an out-of-bounds request reaches the right approval path. Otherwise, a product demo can reward a reassuring refusal while leaving the customer without a completed task.
The economics include a person who learns the business
Pretxt's current pricing page defines a transaction as one model call through its system, including the request and response. It lists paid tiers beginning at $175 a month for Essentials and $2,200 a month for Professional, with different volumes, agent limits and support. Those posted prices describe the offer at launch, rather than evidence of booked revenue or a customer's actual bill.
The same page says every paid plan begins with a forward deployed engineer. Professional includes support through the first production deployment, and Enterprise offers dedicated engineering hours. Pretxt also advertises a Free tier but explicitly says it opens when self-serve setup is ready. Early access currently begins with a conversation about the intended workflow and expected usage.
This creates an operating question for Pretxt alongside the technical one. Customer knowledge must travel from onboarding into a plan that keeps working as the service changes. Each delivery exception, approval path or unsupported claim can expose an assumption that nobody wrote down because an experienced employee already knew it. Engineering involvement gives the company access to that learning, while also making the quality and repeatability of onboarding part of its business.
Related enterprise AI work takes different forms. DevCuration's coverage of Edra's operating-rule approach follows how historical work becomes proposed instructions. Pretxt concentrates on holding an agent to expectations during execution, making the relationship between a written standard and a live action the central product question.
Evidence buyers can bring to the deployment conversation
Pretxt reports that its own testing on Zapier's AutomationBench reduced undesired actions from 18% to zero across two frontier models and 20 support tasks. Its website specifies an October 2026 evaluation and reports about 50 milliseconds of median supervision time compared with roughly 3 seconds for an LLM judge. The comparison is a company-reported result on a defined workload; an independent benchmark's existence does not make every vendor's use of it an independent validation.
A prospective customer can bring its own examples into the early-access discussion. That is where timing, exception handling and coverage become questions about the actual system being deployed. The company's security page says it has completed a SOC 2 audit and describes human approval and history for supervision-plan changes; buyers would still need the relevant report and deployment details for their own review.
The Seed round brings four investment firms behind that customer conversation. Soby and Costello now have to help policy owners turn the standards they already use to manage people into standards an agent can follow at machine speed. The forward deployed engineer arrives where that translation gets specific: beside the customer deciding which promise the business is prepared to keep.
Enterprise AI funding, last 30 days
DevCuration's funding database tracked 18 Enterprise AI rounds totaling $1.5B in disclosed capital over the past 30 days. Recent deals we covered:
- Edra Lands Reported $10M EQT Investment After IT Pilot$10M · Nov 17
- Noeri Raises $110M for Company-Owned, Self-Improving AIFunding · $110M · Nov 17
- Dreamteam Raises $53M for Shared-Context AI Sales CRMSeed and Series A · $53M · Nov 17
- Avanade Agrees to Buy Robiquity for UK AI AutomationNov 14
- Hatz AI Raises $24M Series A for MSP AI BuildersSeries A · $24M · Nov 13
Frequently Asked Questions
How does Pretxt keep a customer request from changing an agent’s authority?
Pretxt describes a supervision plan based on the business’s own expectations. In its illustrative refund walkthrough, a request above the permitted limit is stopped before execution and routed toward the required approval.
Who approves changes when agent activity reveals a gap in the rules?
Pretxt says it can propose refinements based on real activity. The enterprise team reviews and approves changes before those changes take effect.
What do Pretxt’s AutomationBench results establish?
Pretxt reports an October 2026 evaluation across 2 frontier models and 20 support tasks, with undesired actions falling from 18% to zero. These company-run results describe a defined workload and do not establish a guarantee for every production environment.
What should a team prepare for an early-access evaluation?
A team can bring the workflow it wants to delegate, the standards the agent must follow and examples of routine and exceptional requests. Pretxt’s product documentation describes testing supervision plans against examples before deploying a revised plan.
Why is engineering onboarding part of the commercial offer?
Pretxt says paid plans begin with a forward deployed engineer. That support helps translate a customer’s particular use case and operating standards into its first production deployment.
Where the Money Moved
The intelligence briefing of the innovation economy. Funding, M&A, debt and fund closes, read as market signal rather than deal announcements.
Subscribe to Where the Money Moved