Part 1 · Tools, MCP & Sub-agents
The two axes of tools—connectors and MCP—and sub-agents for context isolation.
4. Tools / connectors (+ the MCP industry standard)
▶ Point — Tools are the agent’s hands and feet for taking real action in the outside world. Understand them through two axes: “connectors (Microsoft assets) + MCP (industry standard).” And remember Anthropic’s warning: attaching more tools is not better; it can actually ruin the agent.
4.1 Tools = the agent’s hands and feet
If the loop is “observe → judge,” tools are the channels through which the agent actually acts: sending email, querying a database, creating a ticket, processing an approval, and so on. Without tools, the agent is only a “head that thinks.”
4.2 Anthropic’s three principles for tool design — what makers must remember
Remember: Bloated tool additions are the most common cause of failure.
Effective context engineering addresses tool design directly, and the same guidance applies to the hands-on work in Part 2 (attaching tools).
- Tools must be token-efficient — return concise results
- Capabilities must not overlap — “if even an engineer cannot confidently say which tool to use, the agent certainly cannot do better.”
- Curate the minimum viable toolset — the most common failure is a bloated toolset with too many capabilities and ambiguous choice points
→ Hands-on implication: Do not attach every connector/tool indiscriminately. Use the minimum tools per purpose. If there are ten tools, the agent spends every turn confused about “which one to use.”
4.3 Two axes — connectors and MCP
| Axis | Identity | Nature |
|---|---|---|
| Connectors | 1,000+ Power Platform connectors | Microsoft-owned assets — proven enterprise connections |
| MCP (Model Context Protocol) | Industry standard for agent-tool connections | Not Microsoft-proprietary — connects external tools through a standard protocol |
Why MCP matters: MCP is the specification proposed by Anthropic that became an industry standard for “how agents connect to tools and data.” New Copilot Studio’s support for the MCP server ecosystem (Preview) means that a broad range of tools and apps not reachable through connectors can be attached in a standard way, while still running within Microsoft’s security, permission, and compliance boundaries.
4.4 Inside workflows as well
Workflow nodes can call MCP tools and, when needed, include user approval (human-in-the-loop) before execution (Preview 06/2026).
One-line summary: “Tools = connectors (Microsoft assets) + MCP (industry standard).” Both sit within the same governance boundary. And fewer is better.
5. Sub-agents — connected agents + skills
▶ Point — The essence of sub-agents is “context isolation.” When a specialist agent works deeply in a clean context and returns only a summary, the main agent stays clean. New Copilot Studio moves away from Classic’s child agent model and reconfigures this as “skills + connected agents.”
5.1 Industry principle — why use sub-agents?
Remember: To avoid polluting the main agent’s context.
The main agent handles the high-level plan, while specialist sub-agents focus on work in clean contexts. Each sub-agent may explore deeply using tens of thousands of tokens, but returns only a compressed summary (usually 1,000–2,000 tokens). The detailed exploration context is isolated inside the sub-agent, and the main agent focuses on synthesis and analysis → “separation of concerns.” — Anthropic, Effective context engineering / Multi-agent research system
The planner, generator, evaluator three-agent structure (GAN-style) in the harness design documents is also the standard pattern for role separation. Each agent is designed to cover a specific weakness found in the previous execution.
5.2 Mental model shift in New Copilot Studio
Classic’s child agent concept did not carry over unchanged. Instead:
- Skill = packages a reusable capability
- Connected agents = brings in other agents for collaboration
Why this is better: Child agents ran only inside their own context and were weak at interruptions and intent shifts. Skill-based composition responds much more naturally to interruptions, intent shifts, and multi-intent questions such as “do A and also do B.”
5.3 Current constraints
- At GA, only Copilot Studio agents can connect to each other
- External agent connections based on Microsoft Foundry, M365 Agents SDK, and the A2A protocol are coming soon (for external materials, say only “planned for expansion”)