agent-orchestration
Installation
SKILL.md
agent-orchestration
Core Philosophy
Chaining multiple LLM agents together without deterministic orchestration is a recipe for exponential token costs, cascading hallucination errors, and non-deterministic deadlocks. High-performance agent orchestration treats agents as specialized compute workers within a stateful Directed Acyclic Graph (DAG) or blackboard architecture. Every state transition must be validated against explicit JSON schemas, protected by circuit breakers, and enforce strict token and recursion budgets.
4-Step Agent Orchestration Engineering
Step 1: Topology Selection & State Architecture
- Orchestration Topologies:
- Sequential Pipeline (DAG): Step A -> Step B -> Step C (Ideal for deterministic data extraction, transformation, and publishing).
- Supervisor / Router Pattern: Central orchestrator agent analyzes intent and dispatches tasks to specialized subagents (Research, Code, Review).
- Blackboard / Shared State Architecture: Independent agents read and write asynchronously to a centralized, schema-enforced state store.
- State Immutability & Versioning:
- State must be passed as an explicit immutable snapshot (
AgentState_v1 -> AgentState_v2). - Avoid implicit global variables; every subagent receives strictly the subset of state required for its role.
- State must be passed as an explicit immutable snapshot (