The industry is treating prompts as contracts. They aren’t. They are suggestions.

[Dich01]

Research Report: Reliable Multi-Agent Orchestration & Empirical Hook Behaviors

I’ve documented an empirical investigation into which enforcement mechanisms actually work in Claude Code for multi-agent plugins: exit codes, permissionDecision, bypass vectors, and subagent restrictions.

This is not just a hooks reference. The full report covers:

• Enforcement mechanisms: Which exit codes, permission decisions,
  and hook patterns actually block tool calls — and which ones the
  LLM silently ignores

• Architectural constraints: Why subagents can't orchestrate,
  how to activate a plugin agent as main session, and the namespace
  formats that work vs. fail

• Bypass vectors: How Claude circumvents blocked tools through
  Bash commands and wrapper scripts — and how to close those gaps

• Hook propagation: How hooks inherit to subagents, why this
  causes infinite loops, and the recursion guard patterns that
  prevent them

• Plugin loading gaps: What --plugin-dir actually loads vs.
  what the docs imply, and the bootstrap workarounds required

• End-to-end validation: A complete Express-to-NestJS migration
  with 11 agents, real parallelism, quality gate rejection/correction,
  and 422 tests at 93% coverage

👉 https://github.com/Dich01/claude-code-multi-agent-research