A substantial body of engineering practice now exists for making model-based systems behave. How to constrain them, how to evaluate them, how to catch their failures before those failures reach anyone, where to place a person so that a wrong answer gets stopped while it is still cheap to stop. It is real craft and it is worth having. All of…
Moving an agent from a supervised session into an automated pipeline gets treated as a deployment step. It is a redesign. An agent that behaves well when a person is watching depends on that person in ways that stay invisible until they are gone: someone to answer a clarifying question, to notice when a run stalls, to catch a wrong…
An autonomous consumer of a tool never sees the tool’s internals. It sees the value that comes back, and it reasons over that value as fact. This is the whole premise of tool use: the system asks, the tool answers, the answer becomes ground truth for whatever happens next. Hiding inside that premise is a failure mode that costs nothing to create…
An autonomous agent reads its entire standing instruction layer on every run, and acts on only part of it. That layer is the block of natural-language direction a system carries into every invocation: the persistent instructions behind a served assistant, or the policy prompt that governs an autonomous workflow. Whatever the system, those instructions load on each request, they compete for the…
Every result an automated system produces can be inspected except one. A record can be checked against the source. A number can be recomputed. A summary can be read back against the document it came from. But when a step reports that there is nothing, there is no artifact to examine. The absence of a result and the absence of a…
When a system reads several sources and produces one answer, the sources will sometimes disagree. That is not an edge case to be engineered away. It is the normal condition of any corpus large enough to be worth querying. The design question is not whether disagreement happens but what your pipeline does with it, and there are only two answers: the…
Most teams design a single call and then discover, in production, that they built a conversation. The request arrives underspecified, the model guesses, the guess is wrong in a way nobody notices until the output is downstream, and the fix is a second call that should have been the first. The mistake is not in the prompt. It is in treating the number…
In most systems, results get a schema and failures get a sentence. The success path is versioned, validated, and reviewed; the failure path is a string someone wrote while debugging something else, and it survives untouched into production because nothing that runs regularly depends on its shape. That asymmetry is tolerable when the only consumer of a failure is a human reading…
A synthesis system that reads twenty documents and returns a page of prose has performed a lossy, irreversible transformation, and the loss is not in the words. It is in the correspondence. Somewhere in the generation step, a specific passage in a specific document became a specific sentence in the output, and by the time a reader sees that sentence, the relation…
Every token a language model emits is drawn from a probability distribution over its vocabulary, and the sampling controls are the only place in the stack where you get to alter that draw. They are commonly treated as a quality dial: turn them down when you want the system to be right, turn them up when you want it…