docs: add sub-agent usage guidance to AGENTS.md

Mandates use of agent tool for parallel work and context management.
Encourages batching file reads into 2-3 agent calls instead of
sequential reads.

💘 Generated with Crush

Assisted-by: GLM-5.1 via Crush <crush@charm.land>

AGENTS.md

@@ -10,6 +10,26 @@
 3. **Read docs/PRD.md** - Understand requirements (source of truth)
 4. **Check current state**: `ls -lh output/` and `git log --oneline -10`
 
+### Use Sub-Agents Liberally (MANDATORY)
+
+**You MUST use the agent tool to parallelize work and manage context.**
+
+- **Audit/review tasks**: Dispatch multiple agents in parallel to read different files, test suites, or doc sets simultaneously
+- **Context management**: Use agents for large file reads to keep main context lean. Agents can read, search, and analyze without bloating your context window
+- **Never read 10+ files sequentially**: Batch them into 2-3 agent calls instead
+- **Pattern**: Read 3-4 files yourself for immediate edits, dispatch agents for everything else
+
+**Good patterns:**
+```
+# Parallel audit of different subsystems:
+agent("Read all hooks in config/hooks/live/ and report issues")
+agent("Read all test files and count tests, find stale ones")
+agent("Read all docs and check PRD alignment")
+
+# Context management for large files:
+agent("Read src/security-hardening.sh and report all functions and issues")
+```
+
 ---
 
 ## ⚠️ CRITICAL RULES - READ THESE FIRST