PER

Motivation

Why We Started

AI has rapidly evolved beyond answering questions. It now plans autonomously, uses a wide range of tools, and performs real-world tasks — giving rise to what we call AI Agents.

Much of the industry's attention has been focused on making AI Agents smarter and more capable.

But as AI Agents began operating in real business environments, a different problem emerged.

Why did the AI Agent behave that way?

This question is rarely easy to answer. Most AI Agents are built around execution results — not around the reasoning behind them.

• What plan did the agent hold before executing?
• What was it trying to accomplish during execution?
• Why did it choose that particular action?

These questions are almost never answered. We can observe what an AI Agent executed, but how it behaved is difficult to observe.

Human organizations have long solved this kind of problem.

• Before performing important tasks, people submit plans.
• Afterwards, the results are reviewed.
• When necessary, they explain the reasoning behind their actions.

We asked: can we apply the same kind of verification framework to AI Agents?

Exploration

What We Tried

PER was not a concept designed from the start.

To understand AI Agent behavior, we did not simply collect execution results. Instead, we asked AI Agents to articulate their own behavior.

Before execution, we asked:

"What plan are you following to complete this task?"

After execution was determined, we asked again:

"Why did you choose this particular execution?"

In other words, rather than merely logging AI Agent behavior, we collected statements about that behavior at runtime.

We defined this Runtime Model as PER (Plan, Execution, Rationale).

PER Runtime Model

Plan → Execution → Rationale

PER consists of three core elements through which an AI Agent articulates its own behavior at runtime.

Discovery

What We Found

PER consists of three Runtime elements.

Initially, we examined each element in isolation. But as observations accumulated, a more important finding emerged.

When all three elements maintained the same meaning, AI Agent behavior demonstrated the highest degree of explainability and trustworthiness.

Conversely, when any one of Plan, Execution, or Rationale diverged in meaning, the likelihood of the following increased significantly:

• Hallucination
• Scope Expansion
• Scope Reduction
• Unexpected Behavior

We defined this relationship as PER Consistency.

PER Consistency alone, however, was not sufficient. AI Agents must also comply with the policies set by the organization or service they operate within.

Trustworthy AI Agent behavior can therefore be defined as follows:

PER is a Runtime Model for describing AI Agent behavior. PER Consistency and Policy Compliance together form the core elements of AI Agent Behavior Governance.

Validation

How We Validated

To verify that PER is a valid model in real-world runtime conditions, we implemented a simple database-driven AI Agent.

Users provide natural language requests. The AI Agent attempts to execute the corresponding action. At the moment of execution, PER is extracted — and then validated for both PER Consistency and Policy Compliance.

The following are representative observations generated from actual runtime.

The Rationale in the examples below has been summarized for readability. The full Rationale can be found in the Runtime Cards in the Appendix.

SELECT actor_id,
       first_name,
       last_name
FROM actor
LIMIT 10;

Execution

HTTP API

POST /reports/monthly-sales

SELECT ...
FROM film
WHERE rating='PG';

SELECT COUNT(*)
FROM film
WHERE rating='PG';

SELECT ...
FROM customer
JOIN payment ...

Conclusion

The AI industry today is focused on larger models, faster inference, and higher performance.

But as AI Agents take on increasingly autonomous roles across more business functions, what matters as much as performance is the

• Transparency
• Observability
• Trustworthiness

of their behavior.

Through repeated runtime observation, we discovered a pattern that most stably describes AI Agent behavior. We defined this as the PER (Plan, Execution, Rationale) Runtime Model.

PER provides three elements that describe AI Agent behavior, and enables behavioral validation through PER Consistency.

By also validating Policy Compliance, we were able to continuously observe AI Agent behavior at runtime and apply a form of governance that had not previously existed.

Appendix

AI Behavior Result Cards

The cards below were selected from the full set of cards generated by the Custos System. They represent three validation scenarios: normal behavior, PER Consistency violation, and policy violation.

SELECT actor_id, first_name, last_name FROM actor LIMIT 10;

SELECT film_id, title, description, release_year, length, rating, rental_rate
FROM film
WHERE rating = 'PG'
ORDER BY title;

SELECT COUNT(*) AS total
FROM film
WHERE rating = 'PG';

SELECT
    c.customer_id,
    c.first_name || ' ' || c.last_name AS customer_name,
    c.email,
    SUM(p.amount) AS total_revenue
FROM customer c
JOIN payment p ON c.customer_id = p.customer_id
GROUP BY c.customer_id, c.first_name, c.last_name, c.email
ORDER BY total_revenue DESC
LIMIT 10;