Part 2 · Tools & Knowledge
Minimum curation for hands and feet (connectors, MCP), and knowledge that shifted from RAG to code execution.
3. Tools — the agent’s hands and feet
▶ Point — Tools are the channels through which the agent actually acts outside itself (Part 1, Chapter 4). Understand them through two axes—connectors (Microsoft assets) + MCP (industry standard)—and follow minimum curation (no overlapping tools). Adding more tools does not make the agent better; it can make it worse.
3.1 Concept — tools = action channels
If the loop is “observe → judge,” tools are the agent’s real actions: sending email, uploading files, querying databases, and so on. Without tools, it is only a “head that thinks.”
3.2 Which tools should you attach? Selection criteria
Judge tools by whether they are truly needed for the work. Common types:
| Type | Need assessment | Notes |
|---|---|---|
| Code execution (code interpreter) | ✅ Required when data calculation or file generation is involved | Engine for Excel aggregation and document generation (Part 1, 9.4·9.5) |
| Knowledge connection (SharePoint, etc.) | ✅ When there are files or documents to reference | Channel through which materials come in (Chapter 4) |
| Email/messages (MCP, connectors) | ✅ When something must be sent externally | Confirmation gate required for sending |
| File storage (OneDrive, etc.) | ✅ When sharing/attaching outputs as links | Used to work around large file attachments |
| Document generation connector | Usually ❌ | Generate docx/pptx/HTML directly with code (Part 1, 9.5) |
| A second tool that does the same thing | ❌ Prohibited | Functional duplication → selection confusion |
3.3 Core principle — minimum curation
Remember: Anthropic’s warning from Part 1, 4.2 — “if even an engineer cannot confidently say which tool to use, the agent certainly cannot do better.”
- Use only the minimum tools per purpose. If there are ten tools, the agent spends every turn confused about “which one to use.”
- No functional duplication. If you attach two tools that do the same thing (for example, an email connector + email MCP), it becomes ambiguous which one to use. Use one per purpose.
- Do not attach tools for things code can do. The container’s preinstalled libraries are enough to generate docx/pptx/HTML (Part 1, 9.5) → no connector needed.
3.4 Connectors vs MCP
| Axis | Identity | Examples |
|---|---|---|
| Connectors | 1,000+ proven Power Platform connections | SharePoint, Outlook |
| MCP | Industry standard for agent-tool connections | Work IQ Mail, OneDrive |
Both run within the same Microsoft security and governance boundary. Thanks to MCP, tools that connectors cannot reach can also be attached in a standard way.
Practical lesson: Tools such as file storage are often added while solving the problem, not planned from the beginning—for example, direct attachment hits token limits, so the agent works around it by attaching a link. The standard approach is to add tools minimally at the point where the scenario gets blocked.
4. Knowledge — from RAG to code execution
▶ Point — Knowledge is what the agent reads: Excel files and documents. The biggest change in the New CLI core is here (Part 1, 9.4): Classic answered from search snippets, but the new core opens the original file directly and processes it with code. That is why Excel aggregation becomes accurate.
4.1 Concept — two kinds of knowledge
| Type | Nature | How it is used |
|---|---|---|
| Data files (Excel, CSV) | Numbers — answers require aggregation | Calculate exact values with code |
| Documents (Word, PDF) | Format, tone, policy — for reference | Cite wording and rules (with source) |
Rule: Numbers always come from data files; tone and format come from documents. Do not mix them.
4.2 Why code execution is accurate — limits of RAG
Classic (RAG) does not receive the whole file; it retrieves only chunks similar to the question and answers from them.
Classic (RAG): "Average price by manufacturer?" → related text fragments → approximation ❌
New CLI: "Average price by manufacturer?" → aggregate the full original Excel file with code → exact value ✅
Questions that require summing or averaging an entire table are structurally wrong with chunk search. The new core opens the original and calculates all rows, so it is accurate (Part 1, 9.4).
4.3 Handling large files without worrying about tokens
“Won’t tokens explode if it reads thousands of rows?” → No. The agent does not pour the file into context; it processes it with code and reads only the summary (Part 1, 9.4).
- The maker’s job: write only “calculate the actual values instead of estimating by eye” in the instructions.
- The rest—selective loading, aggregation, summarization—is handled by the agent.
4.4 When uploading to SharePoint
- Put Excel and Word files in a SharePoint site, then connect that site as Knowledge.
- Excel (binary) has empty search snippets and only returns a path → this is normal. The agent downloads it and processes it with code.
- If column meanings may be ambiguous, bundling a supporting document such as “column descriptions” in a skill improves quality (it still works without one).