# Agent System Engineering Agent System Engineering is the discipline of leveraging [[AI Agent Harness|agent harnesses]] to create agentic systems that achieve meaningful goals, efficiently, either in isolation (single agent) or through collaboration (with humans, other agents, or both). Where [[Harness Engineering]] focuses on equipping, guiding, and steering individual AI models with tools, rules, permissions, and skills, agent system engineering operates at the system level: designing how agents are composed, coordinated, and orchestrated to accomplish complex, multi-step objectives. ## Relationship to other disciplines | Discipline | Scope | Focus | |---|---|---| | [[Prompt Engineering]] | Single interaction | How to phrase the question | | [[Context Engineering]] | Information environment | What reaches the model | | [[Harness Engineering]] | Individual agent | How the agent runs | | [[Intent Engineering]] | Goal definition | What must be accomplished | | **Agent System Engineering** | Multi-agent system | How agents work together | Harness engineering is a prerequisite. You need well-harnessed individual agents before you can compose them into reliable systems. Agent system engineering then addresses the emergent complexity of multi-agent collaboration. ## Key concerns - **Agent composition**: which agents exist, what each is responsible for, and how they're specialized (cfr [[AI Agent Identity]]) - **Orchestration**: how agents are coordinated, sequenced, or parallelized (cfr [[AI Agent Orchestration]]) - **Routing**: how tasks are dispatched to the right agent (cfr [[AI Agent Routing]], [[Receptionist AI Design Pattern]]) - **Memory and state**: how knowledge persists across agents and interactions (cfr [[AI Agent Memory]]) - **Communication**: how agents share information, delegate work, and report results (cfr [[AI Subagents]]) - **Panels**: how groups of agents evaluate content from multiple perspectives (cfr [[AI Agent Panels]]) - **Governance**: permissions, audit trails, human-in-the-loop checkpoints ## Architecture patterns - **Single-agent with tools**: one agent orchestrates everything via tool calls and [[Model Context Protocol (MCP)|MCP]] - **Multi-agent pipeline**: agents arranged in sequence, each handling a stage - **Multi-agent swarm**: agents operating in parallel with coordination (cfr [[AI Agent Swarms]]) - **Hierarchical delegation**: a supervisor agent delegates to specialized sub-agents - **Human-in-the-loop**: agents execute within boundaries, escalating to humans when needed ## The microservices analogy The agentic AI field is going through its microservices revolution. Single all-purpose agents are being replaced by orchestrated teams of specialized agents, just as monolithic applications gave way to distributed service architectures. This brings similar challenges: coordination complexity, observability, and the need for clear contracts between components. ## References - ## Related - [[Harness Engineering]] - [[Context Engineering]] - [[Intent Engineering]] - [[AI Agents]] - [[AI Agent Harness]] - [[AI Agent Orchestration]] - [[AI Agent Routing]] - [[AI Agent Identity]] - [[AI Agent Memory]] - [[AI Agent Skills]] - [[AI Agent Panels]] - [[AI Agent Swarms]] - [[AI Subagents]] - [[Receptionist AI Design Pattern]] - [[How coding agents work]] - [[Agentic Engineering]] - [[Levels of AI use]] - [[Agentic Knowledge Management (AKM)]] - [[Claude Code]] - [[SOLID Principles]] - [[Loose Coupling]] - [[Composition over Inheritance]] - [[Software Design Patterns for AI Skills and Agents]]