Simply giving an AI agent a list of tasks is a recipe for misalignment. To get the desired business outcome, you must clearly define what success looks like for its specific role. Without this, the agent will define success on its own terms, often incorrectly.

An AI agent's failure on a complex task like tax preparation isn't due to a lack of intelligence. Instead, it's often blocked by a single, unpredictable "tiny thing," such as misinterpreting two boxes on a W4 form. This highlights that reliability challenges are granular and not always intuitive.

Mozilla's agent worked well because it had a definitive verification signal: a fuzzing build that clearly reports 'you win or you lose'. For projects with more ambiguous outcomes, defining a crisp, automatable success metric is a critical prerequisite for effective agentic work.

Unlike infrastructure where failures are often transient (e.g., network timeout), an AI agent's failure is a persistent reasoning error. Retrying the same flawed logic doesn't fix the problem; it amplifies the negative consequences by repeating the incorrect action with the same confidence and cost.

Treating AI evaluation like a final exam is a mistake. For critical enterprise systems, evaluations should be embedded at every step of an agent's workflow (e.g., after planning, before action). This is akin to unit testing in classic software development and is essential for building trustworthy, production-ready agents.

Agent evaluation is complex because you can't just check the final result. You must also assess the trajectory: did the agent use the correct tools and follow the right process? A correct final answer achieved through a flawed process indicates a brittle and untrustworthy system.

The non-deterministic nature of agentic AI makes traditional pass/fail testing insufficient. Businesses must adopt a multi-dimensional scorecard for every interaction, evaluating metrics like compliance, factual accuracy, latency, and intent recognition, not just task completion.

To ensure model robustness, OpenAI uses a "worst at N" evaluation metric. They sample a model's output multiple times (e.g., 20) on a given problem and measure the performance of the single worst response. This focuses development on eliminating low-quality outliers and ensuring a high floor for safety and consistency, rather than just optimizing for average performance.

While many AI agents produce impressive demos, their real-world utility hinges on reliability. Amazon's Nova Act team argues that for production use cases like UI automation, an agent that works only 60% of the time is effectively useless for business. The critical threshold for value is achieving over 90% reliability, making it the core engineering challenge.