Set up Scout / GitHub Copilot

Tool: VS Code + GitHub Copilot agent mode or GitHub Copilot CLI Source status: Advanced setup

This block gives you a local agent workbench. You will use it to run portable Agent Skills, custom subagents, and MCP tools from the same project folder.

Portability is the point

Agent Skills are plain files. The same folder can run in Cowork, VS Code Copilot, and the GitHub Copilot CLI without rewriting the core instructions.

Steps

Install the GitHub Copilot CLI if you want a terminal runner.
powershell
```
npm install -g @github/copilot
copilot --version
```
Open your project in VS Code.
powershell
```
code .
```
Confirm GitHub Copilot Chat is signed in, then switch the chat mode to Agent.
Org-managed setting
The VS Code setting chat.agent.enabled may be managed by your organization. If agent mode is off and locked, use the Copilot CLI path instead.
Add custom subagents under .github/agents/. The folder name matters. Agent files outside .github/agents/ will not be discovered.
text
```
.github/
  agents/
    researcher.agent.md
    planner.agent.md
    critic.agent.md
```

Give each agent frontmatter. Keep its return shape short so it does not flood the orchestrator.

---
name: capacity
description: Argues from measured capacity and returns a short position.
tools: ['read']
---

You are the capacity subagent.

Return exactly:
POSITION: <one sentence>
BECAUSE: <two sentences max>
RISK: <one sentence>

If one agent must call the others, say so directly in the orchestrator agent.

---
name: arbiter
description: Invokes the council before making a decision.
tools: ['read', 'agent']
agents: ['ambition', 'obligation', 'capacity']
---

You MUST invoke each listed subagent before synthesizing.
Do not simulate their answers yourself.

Run a task from the CLI when you want repeatable, scriptable execution.
powershell
```
copilot -p "Read the local spec and produce the decision memo." --allow-all-tools
```
Approval prompts eat the hack
For a local hack build, --allow-all-tools keeps you out of repeated approval prompts. Use it only in a folder you trust.

For long prompts on Windows, put the prompt in a file and ask Copilot to read it. cmd.exe caps the command line at 8191 characters.

python

from pathlib import Path
import subprocess
import uuid

ROOT = Path(__file__).parent
SCRATCH = ROOT / ".agent-scratch"

def run_copilot(prompt: str) -> str:
    SCRATCH.mkdir(exist_ok=True)
    prompt_file = SCRATCH / f"prompt-{uuid.uuid4().hex[:8]}.md"
    prompt_file.write_text(prompt, encoding="utf-8")
    rel = prompt_file.relative_to(ROOT).as_posix()

    instruction = (
        f"Read `{rel}` in this repository. It contains the task. "
        "Execute it and print only the final answer."
    )
    try:
        result = subprocess.run(
            ["copilot", "-p", instruction, "--allow-all-tools"],
            cwd=str(ROOT), capture_output=True, text=True, timeout=300,
            shell=True, encoding="utf-8", errors="replace",
        )
        return result.stdout.strip()
    finally:
        prompt_file.unlink(missing_ok=True)

Wire MCP tools into VS Code with .vscode/mcp.json.

json

{
  "servers": {
    "local-spec": {
      "command": "python",
      "args": ["mcp_server.py"],
      "cwd": "${workspaceFolder}"
    }
  }
}

Restart the agent session after adding agents or MCP config. Discovery happens at session start.

You'll know it worked when... VS Code can see your .github/agents/*.agent.md files, the CLI can run copilot -p "say hello" --allow-all-tools, and your MCP server appears in the agent tool list.

Apply it to your scenario

Use this setup when your scenario needs more than one perspective. Put the shared facts in files. Put each viewpoint in .github/agents/. Make the orchestrator explicitly delegate, then synthesize.

Set up Scout / GitHub Copilot ​

Steps ​

Apply it to your scenario ​

Set up Scout / GitHub Copilot

Steps

Apply it to your scenario