Building Evidence Based Trust for AI Agents With Vijil

What evidence would convince you that an AI agent is ready to make decisions involving employment, money, healthcare, or legal rights?

In this episode of Tech Talks Daily, I speak with Vin Sharma, founder and CEO of Vijil, about the trust gap preventing many enterprise AI agents from progressing beyond proof of concept.

Vin has spent approximately 30 years building software across security, operating systems, open source, cloud computing, machine learning, and AI. His previous work includes leading engineering at Amazon SageMaker and helping develop 11 AWS AI services.

He argues that AI agents differ from conventional software because they combine autonomy with agency. They can interpret an objective, make decisions under ambiguous conditions, and take action. This raises a deeper question than whether an agent can complete a demonstration successfully: will it remain loyal to the interests of the person or business delegating the task?

Trust is also specific to the job. Vin uses a simple analogy. You may trust a gardener to care for your lawn, but that does not automatically make the same person suitable to babysit your child. An AI agent must therefore be evaluated within the context of its users, task, operating conditions, authority, and potential consequences.

Vin proposes testing three areas. Reliability asks whether the agent can perform its assigned task. Security examines whether it maintains its integrity when facing hostile or noisy conditions. Safety considers what happens when the agent fails and whether the resulting damage remains contained.

This evaluation cannot end when the agent enters production. Models, integrations, data, users, and external conditions change. An agent may drift away from its original purpose, which means businesses need continuous monitoring, testing, and updating across the full AI agent lifecycle.

We discuss how established security practices can be applied to this problem. Trusted execution environments, containment, least privilege, limited-duration access, and bounded models can reduce exposure. Smaller language models may also be better suited to narrow, high-risk tasks than a general model with broad permissions.

Vin offers a three-part framework for governance: personas, purpose, and policy. Personas describe the people and attackers who may interact with the agent. Purpose defines the legitimate task. Policy sets the boundaries between permitted and prohibited behavior.

For high-risk systems, his recommended starting position is that any action not explicitly permitted should be prohibited. A natural-language policy can then be converted into deterministic rules and controls governing the agent's behavior.

Vin's most direct advice concerns evidence. Vibes, demonstrations, and benchmark scores do not prove that an agent is safe for a particular business process. A CISO should expect a complete risk assessment, while a business owner should receive proof that the agent will serve the organization's interests.

His bridge analogy captures the issue perfectly. Engineers do not claim a bridge is safe because it looks impressive during a demonstration. They calculate load, tolerance, failure conditions, and provide test evidence. AI agents acting in consequential workflows deserve a comparable engineering discipline.

If an agent developer asked you to trust their system today, would they be able to provide evidence of reliability, security, safety, loyalty, and contained failure? Listen to the episode and share your thoughts with me.