VS Code — Agent Harnesses (normative extract)

Source: https://code.visualstudio.com/raw/docs/agents/run/agent-harnesses.md (fetched 2026-08-11)

Session targets

Target	Runs where	Code access
Local	VS Code extension host	current workspace
Copilot	Agent Host on your machine	folder or isolated git worktree
Claude	your machine (Claude Agent SDK)	folder or isolated git worktree
Codex	your machine	folder or isolated git worktree
Cloud	provider remote infra	GitHub repo + PR

Why this matters for Patterson

• Claude harness inside VS Code: github.copilot.chat.claudeAgent.enabled. Claude can authenticate/bill through the GitHub Copilot subscription — no separate Anthropic contract needed per seat.
• Claude harness exposes /agents /hooks /memory /init /pr-comments /review /security-review.
• Permission modes: Edit automatically | Request approval | Plan.
• DANGER setting to block via policy: github.copilot.chat.claudeAgent.allowDangerouslySkipPermissions (bypasses ALL permission checks).
• Handoff carries conversation history across harnesses -> a Patterson skill invoked in Copilot can continue in Claude. Argues for skills that are harness-agnostic in wording.
• Worktree isolation = Bypass Approvals, always, non-configurable. Does NOT restrict commands/network. Real containment requires agent sandboxing (/docs/agents/concepts/trust-and-safety.md#agent-sandboxing).
• git.worktreeIncludeFiles needed if agents require gitignored files (.env, deps).
• Cloud third-party agents gated by GitHub org policy: docs.github.com/en/copilot/how-tos/manage-your-account/manage-policies (third-party coding agents toggle).
• Copilot harness limits: only LOCAL MCP servers that require NO authentication. => Patterson MCP servers must run locally + unauthenticated to work in the Copilot harness.