**Lecture 34: Multi-Agent & Biomedical Agents**

**Total Time: 50 minutes**

**Spectrum: Multi-Agent (N×)**
**Code: subagent.py (commit 7f3abb7)**
**Punchline: When 1 agent isn't enough, spawn N.**

**[3 min] 0. Housekeeping**
- Recap Lec 33: built complete single agent (5 files, ~1,200 lines)
- Today: scale from 1 → N agents, then apply to biomedicine

**[5 min] 1. WHAT: Multi-Agent = Agent That Spawns Agents**
- Single agent: 1 conversation, 1 tool set, 1 context window
- Multi-agent: parent spawns children with isolated conversations + specialized tools
- On the spectrum: Multi-Agent (N×) — one step beyond multi-step

**[10 min] 2. WHY: 3 Reasons N > 1**
- **#1 Context fills up** — long tasks exhaust the 200K window. Sub-agent gets a fresh context for each subtask.
- **#2 Specialization** — different tasks need different tools. Explore agent (read-only) vs general agent (all tools). Fewer tools = less confusion.
- **#3 Parallelism** — independent subtasks run simultaneously. "Refactor module A" and "write tests for module B" in parallel.

**[12 min] 3. HOW: subagent.py — 3 Steps**
- **(a) Fork**: create isolated sub-agent with subset of tools + its own conversation
  ```python
  def fork_subagent(prompt, tools, model):
      state = AgentState()  # fresh conversation
      system = build_system_prompt()  # fresh context
      return run(prompt, state, config, system)
  ```
- **(b) Run**: execute agent loop in subprocess/thread, stream events
- **(c) Collect**: gather result, return to parent as tool_result
- **Agent types** (from Claude Code):
  | Type | Tools | Use Case |
  |------|-------|----------|
  | General | All | Default multi-step |
  | Explore | Read, Glob, Grep | Fast codebase search |
  | Plan | Read-only + output | Design before code |

**[15 min] 4. APPLY: Biomedical Multi-Agent Workflows**
- **The self-driving lab** (1 strong case study, not 6 scattered patterns):
  - Planner agent → generates hypothesis
  - Experiment agent → designs protocol, controls instruments
  - Analysis agent → processes data, runs statistics
  - Reasoning agent → interprets results, searches literature
  - Loop back to planner with findings
- **Case study**: mitochondrial DNA mutations in Drosophila (agent_005.png)
  - Human scientist ↔ agents loop
  - Long-term memory: prior experiments inform next hypothesis
  - Literature search: PubMed API as a tool
- **Agent taxonomy for biomedicine** (agent_003.png):
  - Perception (images, genomics, text)
  - Interaction (tool use, multi-agent, human-agent)
  - Reasoning (chain-of-thought, feedback)
  - Memory (long-term, short-term, learning)

**[5 min] 5. Wrap-Up**
- **3 takeaways**: context limits → fork, specialization → tool subsets, parallelism → concurrent agents
- **Code**: subagent.py — fork, run, collect
- **Next Lec 35**: Code Agent — what happens when you scale from 8 tools to 184? (claw-code)