**Lecture 35: Code Agent — Production Architecture**

**Total Time: 50 minutes**

**Spectrum: Code Agent (∞)**
**Code: claw-code (github.com/ultraworkers/claw-code)**
**Punchline: From 8 tools to 184 — what scales, what breaks, what you need to add.**

**[3 min] 0. Housekeeping**
- Recap Lec 34: multi-agent (subagent.py), biomedical agent workflows
- Today: study how a real production agent (Claude Code) is built

**[7 min] 1. WHAT: From nano-claude-code to Claude Code**
- nano-cc: 5 files, 8 tools, 1 agent type, ~1,200 lines — you built this
- Claude Code: 184 tools, 207 commands, 5 agent types, full runtime — what's different?
- The leaked source (Mar 31, 1:23 AM): npm .map file exposed the TypeScript source
- claw-code: community reverse-engineered it into Python (66 files, ~2,138 lines, stubs only)

**[10 min] 2. WHY: 3 Things That Change at Scale**
- **#1 Routing becomes critical** — with 184 tools, the LLM can't see all schemas at once. Need: keyword matching, scoring, deferred tool loading. nano-cc sends all 8; Claude Code routes.
- **#2 Context management becomes engineering** — token budget tracking, transcript compaction (summarize old turns), priority-based truncation. nano-cc: unlimited accumulation. Claude Code: query engine manages every token.
- **#3 Permissions become a system** — not just safe/unsafe flags. Denied tools, denied prefixes, per-tool permission contexts, bash command analysis. nano-cc: 3 modes. Claude Code: full ToolPermissionContext class.

**[15 min] 3. HOW: The Runtime Pipeline (4 stages)**
- **(a) Setup**: platform detection, workspace init, prefetch
  - bootstrap_graph.py → staged startup sequence
- **(b) Route**: score prompt against commands + tools by keyword matching
  - route_prompt(prompt) → ranked list of matching tools/commands
  - This is why 184 tools work — not all loaded at once
- **(c) Execute**: wire up executors, stream events
  - build_execution_registry() → command/tool executors
  - stream_submit_message() → message_start → tool_match → permission_denial → message_delta → message_stop
- **(d) Persist**: save transcript, manage session state
  - persist_session() → disk

**[10 min] 4. HOW: AgentTool/ — Agents That Spawn Agents (16 files)**
- forkSubagent.ts → isolated sub-agent (what we built in subagent.py)
- resumeAgent.ts → pause/continue across sessions (nano-cc can't do this)
- agentMemory.ts → per-agent persistent memory
- spawnMultiAgent.ts → launch N agents in parallel
- 5 built-in types: generalPurpose, explore, plan, claudeCodeGuide, verification
- **Key delta from L34**: resume, persistent memory, custom agent loading from directory

**[5 min] 5. Wrap-Up: nano-cc vs claw-code**
| | nano-cc | claw-code |
|---|---------|-----------|
| Tools | 8 (all loaded) | 184 (routed) |
| Context | accumulate | budget + compact |
| Permissions | 3 modes | full permission system |
| Agents | fork + run | fork + resume + memory |
| Value | working agent | architecture map |

- **Punchline**: nano-cc teaches you how agents work. claw-code teaches you how agents scale.
- **Next Lec 36**: Personal Agent — OpenClaw. Same patterns, completely different domain.