it is binary. a test passes or it fails. code works or it doesn't. you ship when the suite is green.
this assumption is not wrong for deterministic systems. it is structurally inadequate for agentic ones. and the teams that haven't made the epistemological shift yet are building verification infrastructure for a problem that no longer exists.
conventional QA assumes a bounded problem space. you enumerate the inputs. you specify the expected outputs. you write tests that check whether the system maps one to the other correctly. coverage is the metric. green is the signal.
agentic systems don't have bounded problem spaces. an agent operating in a real environment faces a combinatorial explosion of possible states, action sequences, and outcomes that no fixed test suite can enumerate. the number of paths through a non-trivial agentic workflow is not large. it is effectively infinite.
this is not a coverage problem you solve by writing more tests. it is a structural property of the domain. deterministic correctness, does this work, is the wrong question. it has no tractable answer.
the right question is probabilistic: how does this system tend to behave, under what conditions does it fail, and how confident are we in that characterization.
that is a different epistemology. and it requires different infrastructure.
predictive evals measure how a system tends to behave across distributions of inputs. they estimate reliability rather than confirm correctness. this is closer to reliability engineering than QA. reliability engineers don't ask whether a bridge will fail. they model failure probability under load, characterize the conditions that increase risk, and design systems that remain safe within the envelope they've mapped. the goal is not zero failures, it's known failure rates within acceptable bounds, with the ability to detect when behavior is drifting outside those bounds.
predictive evals do the same thing for agents. they produce a behavioral characterization: a map of where the system is reliable, where it degrades, and what the uncertainty looks like at the edges. that map is the foundation everything else is built on, deployment decisions, permission architectures, escalation thresholds.
generative evals don't measure known behavior. they search for unknown failures.
instead of a fixed library of test cases, the eval system generates scenarios dynamically, probing the agent's failure surface, targeting the regions where the predictive model says confidence is lowest, constructing inputs specifically designed to find the failure modes nobody anticipated. it is continuous red-teaming. not measuring what you already know breaks. finding what you don't.
predictive and generative evals are not alternatives. they are complements. predictive evals characterize the reliability envelope. generative evals continuously probe its edges. together they produce something no static test suite can: a live, honest, continuously updated picture of where the agent can be trusted and where it can't.
the systems that combine these two approaches form the backbone of a new infrastructure category.
generative adversarial testing, the continuous function that generates novel failure scenarios faster than the agent can optimize against them. not quarterly red-team exercises. a permanent adversarial function running alongside production.
behavioral drift detection, the instrumentation that tracks whether the agent's failure probabilities are shifting between versions, across input distributions, over time. not aggregate correctness metrics. distributional shift at the level of behavioral tendencies.
trust boundary mapping, the artifact that externalizes the reliability characterization. makes it legible to the humans making deployment decisions. connects it to deployment gates and permission architectures. versions it as the system and its operating environment evolve.
calibrated autonomy controls, the mechanisms that use the reliability map to dynamically adjust what the agent is permitted to do. high confidence in well-characterized environments: more autonomy. approaching the edges of the reliability envelope: escalate, slow down, defer.
none of this exists as an integrated product. the components are scattered across research environments, security tooling, and internal systems at the handful of organizations serious enough to have built them. the integrated stack, generative, predictive, drift-aware, connected to deployment gates, running continuously, is not yet a product anyone is selling.
organizations that think about verification deterministically ask: did the agent pass? organizations that think about it probabilistically ask: what is this agent's reliability profile, and is that profile sufficient for the stakes of this deployment?
the second question is harder. it requires building and maintaining the characterization infrastructure. it requires the specification owner from part one to define what reliability thresholds are acceptable. it requires the adversarial function from part two to keep the characterization honest. it requires the trust calibration layer from part three to translate the characterization into deployment decisions.
probabilistic reliability infrastructure is not a standalone product. it is the connective tissue that makes the rest of the verification stack coherent. without it, you have components. with it, you have a system.
the generation problem produced deterministic abundance. the verification problem is probabilistic by nature.
that shift , from deterministic correctness to probabilistic reliability, is the deepest technical change the AI transition is driving. it is also the least discussed, because it requires admitting that the old epistemology doesn't apply.
the teams that make the shift will build verification infrastructure that actually works. the teams that don't will keep building green test suites for systems that have long since outgrown them.
the infrastructure that answers the right question, not does this work, but how reliably does this work, under what conditions, with what confidence, is the layer the market needs and has not yet built.