FIELD NOTES / QUESTIONS THAT MATTER
Short guides for deciding what to connect, where to place it and how to judge the result.
“Add an AI agent” is too broad to evaluate. “Prepare this maintenance request for dispatcher review” gives the system a purpose, a boundary and an owner. Define what a good output contains, what information is permitted and what should trigger a question.
A remembered answer can outlive its source. Keep provenance, versions and access controls alongside retained context. Decide who can amend it, when it becomes stale and how a corrected record changes future use. More memory is not automatically better context.
Retrieval, network response, tool calls, review and recovery can dominate the work. Compare the complete task rather than choosing a model on a generation-speed claim alone. The useful measure is a dependable, acceptable result with manageable effort.
Do not assume an app or cloud agent will continue when connectivity disappears. Decide what can be available locally, what can be drafted, what must wait and how conflicting updates are resolved. Confirm device protection and retention before enabling local storage.
A general intention to automate does not tell a connector which record it may change. Make the proposed action, affected record and required authority visible. An approval should not silently expand into unrelated actions.
Include incomplete requests, stale sources, contradictory records, unavailable tools and uncertain outcomes. Compare corrections and recovery against the existing process. Expand after reviewing the evidence with the people responsible for the work.
This website presents an architecture and project-enquiry proposition. Its simulator uses prepared rules. Model gateways, retrieval, permissions, integrations and operating support need to be agreed, implemented and tested for each project.