**Lecture 32: LLM APIs & the Agency Spectrum**

**Total Time: 50 minutes**

**[3 min] 0. Housekeeping**
- Roadmap image + reading (smolagents intro, Anthropic API quickstart)

**[15 min] 1. WHAT: From Models to Agents**
- **AI Evolution: The Role of Language**
  - Deep Learning: no LLM, predefined tasks (Lec 11–25)
  - Foundation Model: LLM, on-the-fly tasks via prompts (Lec 26–31)
  - ???: LLM generates text → code!!! (Lec 32–36)
- **Code #1: Control the Loop** — FM: human drives the conversation (VQA). Agent: LLM decides what to do next and when to stop.
- **Code #2: Tool Use** — FM: talks about actions ("you should run rm..."). Agent: actually executes code with side effects.
- **Code #3: Add Feedback** — FM: plans in one shot. Agent: executes, observes errors (e.g., OOM), adapts.
- **Agency Spectrum**: FM → Router (#1) → Tool Call (#2) → Multi-step (#1+2+3) → Multi-Agent (N×) → Code Agent (∞)
  - How much code does the LLM control? More code → more agency.
- **Agents in the Wild**: 3 examples — Coding Agent (Claude Code), Personal Assistant (scheduling), Biomedical Discovery Agent (read papers → hypothesize → experiment → analyze)
- **Demo**: Play first 2 min of "Introducing Claude Code" video

**[8 min] 2. WHY: Three Reasons Agents Matter**
- **Why #1: Unpredictable problems need flexible solutions** — If you can draw a flowchart → hardcode. If not → agent.
- **Why #2: Human-in-the-loop doesn't scale** — VQA: human at every step, 1 task at a time. Agent: one prompt, autonomous, scales to 100 scans.
- **Why #3: Tools multiply intelligence** — LLM alone = brain without hands. LLM + N tools = combinatorial power. Biological parallel: language + tool use = what separates humans from animals.

**[7 min] 3. HOW: Building Agents**
- **What we're building**: Two columns — Coding Agent (Lec 32–35, leaked Claude Code → nano-claude-code) + Personal Agent (Lec 36, OpenClaw)
- **Two "open-sourced" codebases for Claude Code**:
  - nano-claude-code (5 files, ~1,200 lines, working agent) — github.com/SafeRL-Lab/cheetahclaws
  - claw-code (66 files, architectural blueprint, 184 tools) — github.com/ultraworkers/claw-code
- **Roadmap: Follow the Spectrum**:
  - Lec 32: FM (API) → config.py
  - Lec 33: Router + Tool Call + Multi-step → tools.py, agent.py, context.py, nano_claude.py
  - Lec 34: Multi-Agent → subagent.py + biomedical agents
  - Lec 35: Code Agent → claw-code architecture
  - Lec 36: Personal Agent → OpenClaw

**[15 min] 4. Five Things About LLM APIs**
- **(a) API Key: Q → A** — Simplest call: import anthropic, create client, messages.create(). Key params: model, max_tokens, system, messages, temperature, tools (Lec 33). Streaming vs non-streaming.
- **(b) Tokens & Cost** — ~0.75 words/token. Cost = input + output tokens. Model pricing (sonnet $3/$15, opus $15/$75 per M tokens). Context window = memory (200K Claude, 128K GPT-4o, 1M Gemini). Agents fill context fast.
- **(c) Dialog — Messages Format** — Conversation = list of role-tagged messages (system, user, assistant). Append to continue multi-turn. Same format across all providers.
- **(d) Structured Output** — Problem: raw text hard to parse. Solution: prompt for JSON. Even better: tool-based schemas (Lec 33).
- **(e) Vision — Images + Text** — Multimodal: content = list of image + text blocks. base64 or URL. Connects to Lec 30–31 (CLIP, LLaVA). Biomed: histology, radiology, cell counting.
- **(f) In config.py** — Maps a–e to nano-claude-code. config.py (76 lines) handles a + b. Dialog/structured/vision live in agent.py and tools.py (Lec 33).

**[2 min] 5. Wrap-Up**
- **Today**: What (LLM → code → agent), Why (unpredictable, scale, tools), How (API basics a–e)
- **Covered in nano-claude-code**: config.py
- **Next Lec 33**: Tools + Agent Loop — tool schemas, while-loop, permissions, context, full REPL (tools.py, agent.py, context.py, nano_claude.py)