Orchestration Frameworks

Orchestration Frameworks

Building agents from scratch — managing tool calls, context, retries, multi-agent communication — is complex boilerplate. Orchestration frameworks provide abstractions that make it easier to build, test, and deploy agentic systems.

The Main Frameworks

LangChain: the first widely adopted agent framework. Large ecosystem, many integrations (tools, vector stores, memory), comprehensive documentation. Can feel over-engineered for simple use cases — fine for complex pipelines.

LlamaIndex: originally focused on RAG and document retrieval, expanded to full agent support. Excellent for knowledge-intensive applications.

LangGraph: LangChain's graph-based framework for building controllable, stateful agents — explicit nodes and edges instead of an implicit loop. Popular for production multi-agent systems.

Microsoft Agent Framework: Microsoft's unified successor to AutoGen (which has been retired) and Semantic Kernel. Covers both multi-agent conversation patterns and enterprise orchestration.

OpenAI Agents SDK: OpenAI's first-party framework for building agents with handoffs between agents, built-in tracing, and guardrails.

Google ADK (Agent Development Kit): Google's first-party framework for building and deploying multi-agent systems, with tight integration into Gemini and Google Cloud.

CrewAI: high-level multi-agent framework. Agents have roles, goals, and backstories. Accessible for those less familiar with lower-level details.

Claude Agent SDK: Anthropic's own tooling for building Claude-powered agents. First-party, well-maintained.

Bare API + your own code: for simple use cases, a few hundred lines of TypeScript or Python often beats a framework. Frameworks add abstraction but also complexity and version churn.

What Frameworks Provide

Prompt management: templates, versioning, structured outputs.

Tool/function call handling: routing model tool requests to actual functions, returning results, handling errors.

Memory and state: conversation history management, vector store integration, cross-session state.

Agent loops: the observe-reason-act loop, stopping conditions, max iterations.

Multi-agent coordination: spawning sub-agents, passing messages, aggregating results.

Tracing and debugging: recording every step of an agent's execution for review and debugging (LangSmith, Phoenix, etc.)

Deployment helpers: serving agents via API, streaming responses, handling concurrency.

Choosing a Framework (or Not)

A framework makes sense when:

• You're building something complex with many moving parts
• You want pre-built integrations (dozens of tools already built)
• Your team is already familiar with the framework
• You're building quickly and want scaffolding

Skip the framework and use the raw API when:

• Your use case is simple (one agent, a few tools)
• You want full control over exactly what's happening
• You're learning — frameworks hide important details
• Framework abstractions are fighting against what you're trying to do

A common pattern: start with the raw API to understand what you're building, then introduce a framework if complexity warrants it.