Week 1 Assignment

Map a Real Agent's Harness

Pick any production AI agent (Claude Code, Cursor, Manus, Devin, Copilot, etc.) and identify as many of the 9 harness components as you can from its public docs, blog posts, or your own usage.

Deliverables

1. Name the agent and link to its docs
2. For each of the 9 components, write 1-2 sentences on how the agent implements it (or mark it as missing)
3. Identify the weakest component. What would you improve?

Example Format

Agent: Claude Code

Loop: ReAct-style, Think → Tool Call → Observe

Tools: Read, Write, Edit, Bash, Search (~15 tools)

Permissions: Allowlist + bash sandboxing

Context: Auto-compaction at 80% capacity

Memory: CLAUDE.md + session persistence

Errors: Injects errors into context, retries

...