**Lecture 36: Personal Agent — OpenClaw + Course Wrap-Up**

**Total Time: 50 minutes**

**Spectrum: Code Agent applied to personal productivity**
**Code: OpenClaw (github.com/openclaw/openclaw)**
**Punchline: Same architecture, different tools. The agent pattern is domain-independent.**

**[3 min] 0. Housekeeping**
- Recap Lec 35: production architecture (184 tools, routing, context management)
- Today: apply the pattern to a new domain + wrap up Theme 4

**[12 min] 1. WHAT: OpenClaw — Same Pattern, New Domain**
- "The AI that actually does things" — personal agent via WhatsApp/Telegram
- Side-by-side comparison:
  | | Claude Code | OpenClaw |
  |---|------------|----------|
  | Domain | Code | Life |
  | Tools | Read, Write, Bash, Grep | SendEmail, Calendar, BookFlight |
  | Context | Git branch, CLAUDE.md | Contacts, preferences, history |
  | Loop | Same while-loop | Same while-loop |
  | Permissions | "Write this file?" | "Send this email?" |
- **Punchline**: change the tools, keep the architecture. Agent pattern is domain-independent.

**[10 min] 2. WHY: 3 Levels of Agent Safety**
- **#1 Permissions** (what we built in L33): gate individual tool calls. Read=safe, Bash=ask.
- **#2 Sandboxing**: run agent in isolated environment. Docker containers, VM, restricted filesystem. Limits blast radius.
- **#3 Evaluation**: how do you know the agent works? Benchmarks (SWE-bench for coding), sandboxed tests, human review loops.
- **Stakes scale with domain**: wrong code commit < wrong email sent < wrong medical diagnosis. Safety investment must match stakes.

**[10 min] 3. HOW: Building Domain-Specific Agents (3 steps)**
- **(a) Pick your tools**: what actions does your domain need? (PubMed search, DICOM viewer, email sender...)
- **(b) Write your context**: domain-specific system prompt + CLAUDE.md equivalent. Medical agent needs different instructions than coding agent.
- **(c) Set your permissions**: which tools are safe? Which need human approval? Domain determines the answer.
- **That's it.** The loop (agent.py), the dispatcher (tools.py), the REPL (nano_claude.py) stay the same.

**[12 min] 4. The Full Spectrum: L32–L36 Recap**
- **L32**: FM (API) — call an LLM, get text back. config.py.
- **L33**: Multi-step Agent — tools + loop + feedback. 4 files, working agent.
- **L34**: Multi-Agent — fork, run, collect. Biomedical workflows.
- **L35**: Code Agent at Scale — routing, context management, 184 tools.
- **L36**: Personal Agent — same architecture, different tools.
- **The punchline of Theme 4**: An agent is just a while loop + tool dispatch + permission gate. ~1,200 lines. Everything else is scaling.

**[3 min] 5. Wrap-Up & Looking Ahead**
- **Theme 4 complete**: you can build, scale, and deploy agents
- **PS4 due**: agentic systems assignment
- **Final project slides due**: apply agent concepts to your research domain
- **Next**: Final project presentations (Lec 37–39)