Integrating an agent into a real business is rarely hard because of the model; it is hard because of process boundaries. We first break the workflow into describable nodes, then decide which nodes the model may decide and which must keep human confirmation and a reversible record.
Map the process before choosing tools
Discovery answers three questions: where the input comes from, who receives the output, and what happens on failure. Without them, changing models or adding tools only pushes uncertainty downstream.
Keep knowledge, tools and flow separate
We organise source knowledge, external tool calls and orchestration as three separate layers. A business rule change then only touches the flow, a knowledge refresh does not rewrite prompts, and a tool upgrade does not disturb the whole chain.
Observability after launch
Delivery is not the end. Run logs, exception branches and human takeover counts determine whether later tuning is possible. We keep those metrics in the delivery documents so maintenance is measured consistently.
Summary
- Process boundaries before model selection
- Modular organisation supports iteration
- Runtime observation is part of the deliverable
The goal is not to ship every feature at once, but to keep the system maintainable, explainable and evolvable as the business changes.