AI Project · Agent 002
I created a custom agent inside my workplace's approved Microsoft Copilot environment. Its advice draws on enterprise practice and published guidance from major companies, AI labs, Microsoft, and Google. It helps colleagues design, test, and govern AI-supported business processes, from a reusable prompt to a multi-step workflow.
The source material covers workflow architecture, platform constraints, retrieval, file design, testing, permissions, and governance. Source checks resolved conflicts. Nine focused modules organize the findings for retrieval. A three-level experience check adjusts the language and pace. Beginners and experienced builders receive the same knowledge, design method, and quality standard.
Knowledge production
A detailed execution blueprint defined six knowledge responsibilities, source requirements, research checks, and production order. It required the foundation to serve many enterprise processes while using one complex planning workflow as a practical test.
01 · Starting evidence
02 · Research controls
03 · Expert output
Build sequence
Each stage used the output of the previous one. Research established the facts, and the blueprint organized them. The final two stages prepared the technical foundation for retrieval and defined the agent's behavior.
Early access to Microsoft's Cowork platform prompted two focused research runs and the design of an enterprise planning workflow. That practical work exposed the need for file state, handoffs, validation, retrieval, and human approval.
The planning workflow tested the platform research against a real operating process.
The blueprint divided the subject into six knowledge responsibilities, defined the inputs and research checks, and required a foundation that could support any suitable enterprise process.
The information architecture existed before the synthesis model began writing.
Fable, a long-context synthesis model, executed the blueprint in one controlled session. It combined the research, operating context, practical workflow, Google framework, current Microsoft documentation, and contextual answers.
The session produced six technical references, a handbook, and a presentation summary.
The project distilled six expert references into nine shorter retrieval files. Each file covers one topic, uses a descriptive name, and places its decision rule near the top.
The new structure made the technical method retrievable during a business conversation.
A separate instruction contract added the experience check, process questions, right-sizing rules, one-step delivery, testing requirements, and governance boundaries.
The completed advisor produces prompts, file designs, tests, and concrete next actions from the knowledge library.
Knowledge architecture
The six foundation documents serve an expert reader. The agent uses a second layer of nine shorter documents. Each covers one topic, uses a descriptive filename, and places its decision rule near the top. A separate instruction prompt controls the conversation.
This separation supports targeted maintenance: one knowledge topic can change without rewriting the agent's behavior, while conversation changes leave the source foundation intact.
Inside the agent
The nine modules form the agent's RAG knowledge layer. They help the agent understand the problem, design the solution, and control the result. Each module supports a defined decision.
Advisory
Build
Assurance
Help design
The instruction contract gives business users access to the foundation through a short conversation and keeps the work tied to their process.
First message
The opening confirms the request, checks the user's experience, and asks two to four questions about the process.
Solution size
A one-off task may receive one prompt. Recurring multi-file work may justify an agentic workflow.
Division of work
The agent writes prompts, proposes file changes, and diagnoses failed tests. The user provides context, runs the step, and reviews the result.
Control
The agent adds counts, known-answer cases, samples, permission checks, and human approval before live use.
Closing principle
A reliable result depends on the system around the model. Clear task boundaries and curated knowledge define the work. Structured files preserve state between steps. Checks and approval points expose errors before they affect a decision. Together, these elements give people an output they can test, review, and improve.