agent-orchestrationlisted

**CONSUMES**: User multi-agent / orchestration task + target workflow description + step count / duration estimate + optional existing framework choice **PRODUCES**: Applied orchestration judgment rules + framework selection rationale + topology decision (supervisor/swarm) + durability/checkpoint plan + HITL interrupt design + tool-permission audit # Agent Orchestration Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 **License**: Apache 2.0 --- ## What This Pack Does AI agents wire up multi-agent systems by copying a framework's quickstart. They pick LangGraph because it is popular, not because the workflow needs deterministic state. They build swarms with 8 peer agents and never reason about the O(n²) failure surface. They run 300-step agents on bare retry loops, then are surprised when step 280 crashes and restarts from step 1 — re-sending the emails it already sent. They auto-approve every tool call because adding a human gate "later" never happens. This pack embeds the judgment rules that orchestration engineers apply automatically — rules from framework documentation, Temporal durable-execution patterns, and production reliability data. **Pack = orchestration judgment. Your workflow system = process constraints. No overlap.** --- ## Cross-Cutting Rule: The Complexity Cliff — Reliability Decays Exponentially with Step Count > **Cumulative agent reliability is `P(fail) = 1 - (1 - p)^s`, where `p` is