Show HN: A practical framework for evaluating agentic AI systems

In classical ML, verification usually meant checking whether the prediction matched a label or whether a score was close enough to a reference. For generative models, it often means comparing the output text with a ground truth.

An agent can also have ground truth, including ground truth for the path it took. But creating it requires subject-matter expertise, and it rarely takes the form of one canonical answer. Several different trajectories may be equally correct. The practical alternative is to describe what acceptable behavior looks like and what the agent must achieve.

Agentic eval has to score the agent’s entire behavior instead of a single output, and there’s no equivalent reference to check it against: a resolution that might take five turns, three tool calls, and a fallback to get there. So now you have to care about far more than the final answer. A 100% completion rate says nothing on its own about how the agent got there. The agent can call the wrong tool and self-correct, retry blindly until something sticks, or wander through a detour that happens to land on the right final state anyway. Trusting the agent means checking the path it took:

Three runs of the same prompt, three different paths:

So guard the eval case with several checks instead of one score, and write each one against something that should happen: this tool was called, with arguments that came from the conversation rather than from thin air, before that one, and the whole thing finished inside your time and cost threshold. Each check is narrow on purpose, so a run that lands on the right answer the wrong way fails at least one of them.

What you don’t do is pin the whole path. Depending on the use case, the agent may need room to explore, and two runs that reach the same place by different routes can both be fine. Assert the steps that must be there and the boundaries it must stay inside; leave the rest of the trajectory free.

A score earns trust when it behaves predictably: full marks mean the job was done perfectly, zero means it was not done, and better work receives a higher score. In agentic systems, the challenge is that this number often comes from an LLM judge rather than a simple function.

The same properties apply to classical and agentic metrics: they should be bounded, and monotone. These are not new requirements. Boundedness and monotonicity are established principles in metric design and measurement theory (Amigó & Mizzaro, 2020). They matter just as much when the score is produced by an LLM.

Prove each one before you trust a number: which cases actually produce 100%, which produce 0%, and confirmation that the ranking in between holds. That part hasn’t changed. At deepsense.ai we went one step further, because an average doesn’t give you the whole picture. An agent that returns a solid 70% on every run and an agent that swings between 0%, 30%, 70% and 100% can average out to the same place — but in production you need the first one. Reliability is the property you’re buying, so we built it into the metric rather than checking for it afterwards. It’s called the Business Utility score – it penalizes high variance across runs instead of averaging it away, and the write-up is worth a read if you’re designing your own.

What changed is how you get there. §3 covered the first half: the object being scored moved from a response to a task. The second half is the scoring mechanism itself. Because an agent can reach a good outcome by many different paths, there’s rarely one canonical output left to exact-match against, so rubrics, state checks, and LLM judges take over from exact match. That’s convenient and dangerous in the same move: the judge is now the thing that has to carry boundedness, anchoring, and monotonicity, and an LLM judge doesn’t earn them just by sounding confident.

The same care applies to how many metrics you keep. Three scores you have validated and know how to act on are worth more than fifteen nobody has checked against the properties above. If a number moving wouldn’t change what you do next, it is costing attention and returning nothing. That is separate from the narrow checks feeding each score, which should stay numerous. It’s the headline numbers you report and gate releases on that should be few.