Part 1 · Claude Code Comparison & References
Same harness philosophy, different purpose — plus improvements over Classic and reference materials.
10. Claude Code vs Studio CLI agent
▶ Point — Both are products of the same harness philosophy (Chapter 1). The difference is purpose: Claude Code maximizes developer autonomous execution, while the Studio CLI agent is an “enterprise harness” that maximizes M365 data integration + guaranteed human intervention. This chapter summarizes (1) their shared harness, (2) the decisive differences, and (3) improvements over Classic Studio.
10.1 Both are harnesses — element comparison
Comparing Claude Code’s five harness elements (Tool Loop, Sandbox, Persistence, Observation, Interruption) with the current CLI agent:
| Harness element | Studio CLI agent | Evidence |
|---|---|---|
| Tool Loop | ✅ | Repeated calls to bash, grep, view, edit, and python |
| Sandbox isolation | ✅ | Azure Linux container, sandbox user, /dracarys/ blocked |
| Persistence | ✅ partial | Maintained within a session through session-store.db; cross-session depends on M365 |
| Observation | ✅ | Tool results feed back into context (including preToolUse hooks) |
| Human-in-the-loop | ✅ strong | User confirmation for every response; no unauthorized automatic execution |
| Skills/specialization | ✅ | /app/skills/ — specialized skill modules separated |
| Memory layers | ✅ | Short-term (DB) + long-term (M365 cloud) separated |
→ Conclusion: this is indeed “an agent based on harness engineering.”
10.2 Decisive differences — same philosophy, different targets, purposes, and governance
Remember: It is a misconception that the Studio CLI agent is a tool “only for HITL, where a person must intervene every time.” When connected to workflows (Chapter 7), it becomes autonomous automation as-is. So the difference is not whether automation is possible, but who uses it (target), for what (purpose), and under what controls (governance).
The two harnesses share the same philosophy: isolated sandbox + tool loop + observation and iteration. They diverge along three axes.
| Comparison axis | Claude Code harness | Studio CLI agent |
|---|---|---|
| Target — who uses it | Developer-centered | General users and semi-developers (business users) |
| Purpose — what for | Maximize local, project-level development automation | Organization-level deployment and frontline work improvement |
| Governance and control | Relatively free — local-centered work is possible; enterprises must provide controls separately | Only inside enterprise-controlled infrastructure — works with centrally deployed resources (knowledge, data, tools) |
| Automation method | Agent executes autonomously across multiple steps and judges completion itself | Default is human intervention (HITL); autonomous execution when connected to workflows (Chapter 7) — choose intervention or automation |
| Inter-agent communication | Multi-agent orchestration | Expands through connected agents (Chapter 5) |
| Data and development integration | Git-centered work | M365 ecosystem-centered (+ for developers, Git CI/CD and VS Code Extension support instructions and skill development) |
Claude Code harness → developers · free autonomous local execution (Agentic)
Studio CLI agent → business users · data integration inside enterprise governance + optional automation (Enterprise)
One-line summary: The same harness philosophy is implemented for different targets, purposes, and governance models. Claude Code lets developers execute autonomously and freely on their local projects; Studio CLI lets business users connect to data inside enterprise-controlled infrastructure and automate with workflows when needed. This is not a superiority contest between HITL and automation; it is a difference in who uses it, for what, and under what controls.
10.3 Improvements over existing (Classic) Studio agents
They share the “Copilot Studio” name, but the execution model itself is different. Improvements visible from the runtime evidence in Chapter 9:
| Area | Classic Studio | New Studio CLI agent | Improvement |
|---|---|---|---|
| Execution model | Topic tree (conversation flow design) | Isolated container + agentic loop | Multi-step and long-running work; branch explosion resolved |
| Knowledge handling | Answers based on search snippets | Real-file download + Python large-scale processing | Large files +8.3, code interpreter +41.8 (NDA) |
| Outputs | Mostly text responses | Rich files such as Word/PPT/Excel/PDF (created/) |
Dramatically broader output formats |
| Tool execution | Calls predesigned actions | Real-time tool loop with bash, python, and grep |
Adaptation, retry, recursive execution |
| Code execution | None/limited | Preinstalled Python runtime (isolated) | Large-scale data analysis |
| Memory | Variables and topic state | Short-term session-store.db + long-term M365 |
Context tracking and checkpoints |
| Security boundary | Platform governance | Per-session isolated container + sandbox user + engine blocking | Session isolation and least privilege |
FAQ response: “Isn’t it just an old bot with the same name?” → No. Classic was a tool for drawing conversation flows, while the New CLI agent is a runtime that runs real tools, code, and files inside an isolated container for each session. The “AI core rebuild” from Chapter 0 is evidenced by the container in Chapter 9.
References
Industry (public) — primary sources
- Anthropic Engineering, Effective context engineering for AI agents — attention budget, context rot, just-in-time, compaction, note-taking, sub-agents, tool design
- Anthropic Engineering, Effective harnesses for long-running agents — initializer + coding agent, feature list, incremental progress, clean state, browser E2E validation
- Anthropic Engineering, Designing harnesses for long-running apps — planner, generator, evaluator (GAN-style), “start from the simplest solution”
- Anthropic, Building Effective Agents / Multi-agent research system / Writing tools for AI agents — supplementary
- Anthropic Learn (anthropic.com/learn) — learning course
- Examples: Claude Code, GitHub Copilot CLI, OpenAI Codex CLI