v1.0.0 - Field report

First public release of the field report.

A candid, dated field report on building a multi-country document-compliance agent in full-AI mode on Copilot Studio (snapshot: August 2026, GPT-5.5 baseline, GPT-5.6 compared).

What's inside

• Architecture — an agent flow that owns pagination + five single-purpose agents.
• Execution walkthrough — one full loop, step by step, with observed timings.
• Data contracts & the deterministic output pattern — "the model judges, scripts build".
• Harness limits (flagship chapter) — the constraint -> decision table, including "agent laziness".
• Model selection — GPT-5.5 vs GPT-5.6: quality vs cost & latency.
• Where to remove AI in production — the de-AI map.
• Safeguards — the "never do this" list.
• Reproduce it — table structure + wiring, no client data.

Notes

• Client-neutral: all data, documents, tenant and organization details removed; the running example is fictional.
• The two analyst skill packages are intentionally not shipped (client-derived).
• Not affiliated with or endorsed by Microsoft. Product names belong to their owners.