Tech leaders face a dilemma as they prepare to put AI agents into production. They want the higher productivity and lower costs that AI agents can provide. But they fear unknown risks from AI capabilities that differ from the rest of their software stack. When an agent makes decisions and takes actions on its own, it might be doing its best work. It might also be doing its worst.
As AI agents become more capable, organizations want assurance that their oversight can keep up. How do they strike the right balance between innovation and risk? It comes down to how well they can assess, measure, and maintain trust in their agents. The more trust agents earn, the more autonomy they can be granted.
Trust as a line of credit
Think about the way credit card risk is managed. When you apply for a card, the issuer assesses your credit history and ability to repay before setting a credit limit. Then, once you have your card, payments are subject to real-time authorization and fraud controls. Unusual transactions can trigger additional scrutiny. Paying your card on time may improve your credit score, but it won’t automatically increase your credit limit.
AI agents should be treated similarly.
Test an AI agent before production for resilience, reliability, and performance. Assign it a “trust score” based on how well it performs. Once the agent is deployed, keep the checks coming. Subject each consequential action to real-time policy enforcement, and monitor for anomalies. Treat evidence of improved agent performance as a reason to keep the agent operating, not necessarily to grant it more autonomy or scope within the organization’s risk limits.
See where your agent stands today: sign up for a free trial to evaluate its trustworthiness at ivijil.ai/benchmark
Now imagine the card could make purchases on its own. You’d want rules that keep it within budget, and evidence that it spends the way you would. That’s an AI agent.
Earning that trust requires a DevOps-style loop:
- Test intensively before launch
- Ship with limits (based on the trust score)
- Monitor the agent in production.
- Address identified points of failure, retest to assure trust, and redeploy the agent into production.
The loop never ends. Agents’ behavior can change over time. Continuous testing and monitoring are necessary to spot areas where their trustworthiness has declined and respond with appropriate fixes.
Agents gone feral
Once you’ve trained your agent and it’s doing the right things, it might appear domesticated. But appearances can be deceiving. There are several ways an agent can go feral after launch:
- Drift: Changes in the environment mean the old rules no longer apply. Agents may behave unexpectedly in a new environment.
- Malicious actors: Jailbreaks and prompt injections can cause agents to violate policies or misuse tools that they are authorized to access.
- Ungoverned agents: Legitimate agents, but operating without adequate visibility, accountability, or a formal evaluation lifecycle.
- Multi-agent interactions: Individually permitted actions can combine into outcomes that violate organizational policy, even without deliberate rule-breaking.
Even well-tested systems can fail in production. With agentic systems, changes to models, prompts, tools, or workflows can alter behavior after pre-deployment testing. Without ongoing monitoring and testing, feral behavior can creep in.
Once you’ve trained your agent and it’s doing the right things, it might appear domesticated. But appearances can be deceiving.
The fix is to establish a baseline and enforce governance for each agent and its capabilities. Give the agent an identity, test it, and continually score its trustworthiness. The goal is an agent that acts in the principal’s interests, is accountable to the principal, and isn’t vulnerable to going feral.
To that end, guardrails are critical, but they shouldn’t be the primary way to keep agents in line. This isn’t Mario Kart, where bouncing off the walls is part of the game.
Ideally, agents should stay on their tracks, like trains. In this case, the tracks are policies that describe the agent’s purpose and the organizational requirements it must meet. Vijil’s approach builds on the rails analogy, extending it across the agentic lifecycle by translating policies into evaluation criteria and runtime controls.
Fiduciary: the destination
Trust is easy to define. But it’s hard to program into an agent. So we borrow a standard from finance: the fiduciary. A fiduciary owes duties of care and loyalty within the scope of the relationship.
Trust is easy to define. But it’s hard to program into an agent.
Under U.S. law, an investment adviser acting as a fiduciary must serve the client’s best interests and must not put its own interests ahead of the client’s.
We use fiduciary as an analogy for the behavior we want from a trusted agent. For AI agents, fiduciary describes an intended behavioral standard, not legal fiduciary status or a guarantee of loyalty. The aim is to serve the organization’s interests rather than the model vendor’s. But agents are different from humans. They act without intent, and (despite how easy it is to anthropomorphize them) they lack personhood. So for agents, being a fiduciary is a state of behavior. It’s an objective standard that can be judged by what the agent does.
The goal is to move your agents from feral to fiduciary. Do it as quickly and efficiently as possible. Vijil customers have cut time to trust by 4x.
Building trust into the agentic lifecycle
With Vijil, agents pass through the same continuous evaluation loop as code in a DevOps pipeline: build, ship, run, improve. The different phases of this loop verify the agent’s reliability, security, and safety. They measure performance and fix points of failure. They detect when trustworthiness improves or deteriorates. Teams that need data to stay in-house can deploy on-premises, in their own cloud, or in an air-gapped environment. The hosted free tier sends evaluation traffic and decision traces to Vijil.
Every agent you deploy is serving someone’s interests. The only question is whether that interest is yours. The path from feral to fiduciary starts with evidence, not faith.
The post From feral to fiduciary: making AI agents serve your interests appeared first on The New Stack.