Key facts
- Comparison date
- July 13, 2026.
- Pick after
- The agent spec names tools, state, and evals.
- Default option
- Custom code is valid for narrow loops.
- Fast check
- Run the framework chooser.
Comparison date: July 13, 2026
Agent frameworks change quickly. This comparison is a current implementation aid, not a permanent scoreboard. The goal is to match architecture to project shape: explicit state machines, collaborative role systems, provider-native SDKs, codebase agents, or plain application code.
Durable execution, streaming, human-in-the-loop, and persistence.
The practical comparison
| Option | Best fit | Strength | Watch out |
|---|---|---|---|
| LangGraph | Stateful, long-running, branching workflows where the graph matters. | Explicit graph model, persistence, streaming, memory, and human review patterns. | Extra structure can feel heavy for a small single-agent task. |
| CrewAI | Role-based multi-agent systems with crews, tasks, and flows. | Readable mental model for agents collaborating on business workflows. | Role names can become theater unless tasks, tools, and state are concrete. |
| OpenAI Agents SDK | Python or TypeScript agent apps with handoffs, guardrails, tools, sessions, and tracing. | Small set of primitives, OpenAI-hosted tools, sandbox agents, tracing, and provider adapters. | Tightest fit when your stack is already OpenAI-centered. |
| Claude Agent SDK | Self-hosted tool-rich agents needing permissions, hooks, MCP, subagents, and resumable sessions. | Python and TypeScript SDKs with built-in filesystem, shell, search, and agent-loop capabilities. | You operate the process, sandbox boundary, and session storage; use only the powerful tools the task needs. |
| Claude Managed Agents (beta) | Long-running or asynchronous Claude agents where managed sessions and sandboxes reduce infrastructure work. | Managed tool execution, persistent state, cloud or self-hosted environments, and versioned agent definitions. | Beta surface; server-side session state has different data-retention and compliance implications. |
| Custom orchestration | Narrow products with a fixed loop and limited tools. | Maximum clarity, no framework dependency, direct control over state and tests. | You must build tracing, retries, approvals, and state discipline yourself. |
LangGraph
Choose LangGraph when the agent is really a stateful workflow with branching, interruptions, and recovery. Its graph model makes the control flow explicit, and its persistence patterns are useful for resumable work, human-in-the-loop checkpoints, and long-running tasks. If the product needs a durable thread that can pause, resume, stream progress, and recover after tool failures, LangGraph deserves serious consideration.
The tradeoff is complexity. If your "agent" is a single tool call followed by a final answer, a graph may be ceremony. Use LangGraph because you need state and control, not because the word graph sounds production-grade.
CrewAI
CrewAI is strongest when the product team naturally thinks in roles: researcher, analyst, writer, reviewer, operator. Its concepts of agents, crews, tasks, flows, memory, knowledge, and observability can make multi-agent collaboration legible. Flows are important because they keep orchestration structured instead of leaving every decision to the agents.
The risk is performative multi-agent design. Do not create three agents just because a workflow has three paragraphs. Create multiple agents only when the roles need different instructions, tools, permissions, or evaluation criteria.
Design agents, orchestrate crews, and automate flows.
OpenAI Agents SDK and Claude Agent SDK
Provider SDKs are attractive when they line up with your product surface. OpenAI's Agents SDK gives Python developers agents, handoffs, sessions, hosted tools, guardrails, tracing, and sandbox agents in one stack. That is a compact path for teams already building on OpenAI APIs and wanting provider-supported observability.
Claude Agent SDK is a different fit. It comes from the Claude Code agent surface and emphasizes built-in tools, permissions, hooks, MCP, subagents, session management, and code execution contexts in Python and TypeScript. It is useful beyond coding, but its broad tool harness should earn its place over a small custom loop.
Anthropic now also offers Claude Managed Agents in beta. That is a deployment choice, not a drop-in synonym for the self-hosted Agent SDK: Anthropic runs the agent harness and sandbox, stores stateful sessions, and targets long-running asynchronous work. Evaluate its retention and compliance boundaries before choosing it.
Custom orchestration
Custom code remains a serious option. A product with one model call, two tools, one approval gate, and ten evaluation fixtures may be clearer as normal application code. You can still use provider tool calling, structured outputs, queues, databases, and traces without adopting a full framework.
Verdict
Use LangGraph for explicit stateful orchestration. Use CrewAI for concrete role-based collaboration. Use OpenAI Agents SDK for OpenAI-centered Python or TypeScript agents with first-class tracing and guardrails. Use Claude Agent SDK for self-hosted, tool-rich agents; evaluate Claude Managed Agents when managed long-running sessions fit the risk model. Use custom orchestration when your product is smaller than the framework.
Sources used
- LangGraph OverviewLangChain Docs. Accessed 2026-07-13.
Positions LangGraph as an orchestration runtime for durable execution, streaming, HITL, and persistence.
- CrewAI DocumentationCrewAI. Accessed 2026-07-13.
Introduces CrewAI agents, crews, flows, guardrails, memory, knowledge, and observability.
- OpenAI Agents SDKOpenAI. Accessed 2026-07-13.
Documents the Python Agents SDK primitives: agents, handoffs, guardrails, tracing, sessions, tools, and sandbox agents.
- Agent SDK OverviewClaude Code Docs. Accessed 2026-07-13.
Documents Claude Agent SDK built-in tools, hooks, subagents, MCP, permissions, sessions, and production use cases.
- Claude Managed Agents OverviewClaude Platform Docs. Accessed 2026-07-13.
Documents Anthropic-managed, stateful agent sessions and sandboxes for long-running and asynchronous work; currently beta.
- GuardrailsOpenAI Agents SDK. Accessed 2026-07-13.
Explains input, output, and tool guardrail boundaries in OpenAI Agents SDK workflows.