Insights
What must an agent specification contain to be buildable?
DIRECT ANSWER
An agentic process specification is a description from which an AI agent can be built without guessing. A buildable specification contains five things: the structure (a process decomposes into workflows, workflows into agents, agents into technical tasks), the data flows (where each step's input comes from and where its output goes), the acceptance conditions (by what criteria a step counts as done), the human control points (where a human reviews, approves or decides) and the risk classification (what is allowed to go wrong in a step, and what is not). In most failed AI projects the problem is not the model but the fact that at most one of these five has been defined.
Structure: an agent is small, the process is the whole
The most common mistake is to specify "an agent that handles customer service". That is not a specification; it is a wish. A working specification decomposes the whole into parts small enough that each agent has a few clear technical tasks whose success can be verified. The whole emerges from connecting small, verifiable parts into a process, not from one large agent attempting everything.
Not every step is AI, nor should it be
In a well-specified agentic process, a significant share of the steps are deterministic tools: retrievals, transformations, checks and transfers that behave the same way every time. A language model is used where interpretation is needed. This division is among the most important decisions in the specification: it determines the process's predictability, its costs and the traceability of its errors.
Acceptance conditions make "done" verifiable
For every step, the specification defines how you know the result is acceptable: the format, the content requirements and approved example cases against which the output is compared. Without acceptance conditions, "does it work" is an opinion. With them, it is a checkable claim.
Human control points belong in the specification, not in a retrofit
Where a human approves before the process continues, which cases always route to a human, and who has the authority to stop: these decisions are made at specification time and written into the structure. Oversight added afterwards is typically both too heavy and in the wrong places.
Frequently asked questions
Q:Who writes the specification?
The people who run the process describe the work; we structure it into a specification. Neither succeeds alone: the practitioners have the knowledge, we have the structure and the method.
Q:How precise does it need to be?
Precise enough that two different implementers would build essentially the same process from it. That is a practical test of a specification's readiness.
Q:Does the specification go stale?
Not if changes are made through it and the implementation is regularly compared against it. That is exactly what our governance service exists for.
Q:How does this differ from a traditional process description?
A traditional description is documentation for people to read. An agentic specification is machine-readable, and both the human-facing documents and the implementation baseline are generated from the same source.
Next → 04 // Verifiability as competitive advantage