LLM Output Is a Measurement. Sensor Fusion Solved This.

LLM output is a measurement with uncertainty, not an answer. If you have worked with sensor fusion, that is not a metaphor.

I spent years in autonomous driving and sensor fusion. Now I work on agent systems and keep running into the same architecture — rebuilt from scratch every time, without the pieces that made it work.

In sensor fusion, everything starts with the prior — a prediction from what the system already knows, before the sensor speaks. In agent stacks, that same prior is a context layer: customer state, entity relationships, business semantics, assembled before the LLM reasons rather than as a static prompt. Without it, the model has no grounding and the reasoning drifts quietly.

A prior alone does not tell you how much to trust what comes next. The LLM is a powerful sensor, but noise compounds: 95% accuracy per step becomes 60% over ten. That is not only a model problem. That is a pipeline that trusts every reading equally.

Sometimes the output is semantically correct but structurally invalid. That is not a model failure. I have seen this before — a sensor reading in its own coordinate system, never converted into the shared frame the pipeline trusts. In practice: an agent returns a tariff code the billing system does not recognize. More often a coordinate mismatch than a hallucination.

So you validate. Many pipelines already validate, repair, and retry — but without a calibrated loop. The validator does not compute what failed or relative to what the system expected. In sensor fusion, the innovation gate measures deviation against expected variance. Most agent pipelines skip that: no structured error, just a binary decision. That is why retries keep failing. The fix is not a better retry. It is computing the error first — feeding the model what failed and why.

Most agent failures are not AI failures. They are integration failures — business logic, model instructions, and validation rules never calibrated against each other. You cannot gate what you have not defined. The fix starts before the LLM: make the business contract explicit first.

Alignment alone is not enough. When something breaks in production, you need to reconstruct exactly what the system knew and decided — every prior, every output, every gate decision. Without that trace, you are debugging a black box.

The complete pipeline is not new, but it has to be whole: prior, measurement, gate, update, log, recovery. The accepted output updates the prior — that is how an agent accumulates understanding across calls. Without a trust policy, it quietly accumulates error instead. Remove any piece and the system is not ready for production. If it cannot recover from a failed step, it is not a system. It is an expensive demo.

Recovery before capability. That was the order in autonomous systems. Invisible in a demo. The only thing that matters in production.