Four layers are built once and kept up to date. LLMs help at build time by proposing content, and people approve it. At run time, only one step uses an LLM.
The LLM understands the question. Everything else is governed code.
The same agent answers the same five questions over the same data products. Modes 2 to 4 add each context layer by handing it to the agent whole. Mode 5 adds the enterprise context factory, which builds a small governed context pack for each question.
✓ correct · ✗ wrong · – declined. Context sizes for whole layers are illustrative, at enterprise scale.
Sixty requests from four teams flow through the agents. Switch on the harness, then the mesh, and watch the risks disappear and the work start to be shared.
Each mode replays the same 60 requests, so the comparison is fair. Cost figures are illustrative.