Overview
| Aspect | LangChain Deep Agents | Claude Agent SDK | Codex SDK |
|---|---|---|---|
| Use cases | Custom general-purpose agents (including coding) | Custom AI coding agents | Prebuilt coding agent that can execute coding tasks |
| Model support | Flexible and model-agnostic (Anthropic, OpenAI, and 100s others) | Tightly integrated with Claude models (Anthropic, Azure, Vertex AI, AWS Bedrock) | Tightly integrated with OpenAI models (GPT-5.3-Codex and variants) |
| Architecture | Python SDK, TypeScript SDK, and CLI | Python SDK, TypeScript SDK | TypeScript SDK, CLI, desktop app, IDE extension, cloud interface |
| Execution environment | Local, remote sandboxes, virtual filesystem | Local | Local, cloud |
| Deployment | LangGraph Platform | Self-hosted | N/A |
| Frontend | Integration with React | Server-side only | Server-side only |
| Observability | LangSmith tracing & evaluations | N/A | OpenAI traces & command-line evaluations |
| Security configurability | Composable, per-tool human-in-the-loop | Permission system with modes, rules and hooks | Built-in tiers using approval modes and OS-level sandboxes |
| License | MIT | MIT (underlying Claude Code is proprietary) | Apache-2.0 |
Key differences
LangChain Deep Agents:- Model flexibility: Swap model providers (Anthropic, OpenAI, or 100+ others) at any time and run evaluations.
- Long-term memory: Persist context across sessions and threads with the Memory Store
- Sandbox-as-tool pattern: Run individual operations in isolated sandboxes from different providers while the agent runs outside, or run the full agent inside a sandbox
- Virtual filesystem: Use pluggable backends (in-memory, disk, durable stores, sandboxes) for context and checkpoint management
- Production deployment: Deploy via LangSmith or self-host with the Agent Server
- Observability: Use LangSmith for native tracing and debugging
- Standardize on Claude: First-class support for Claude models across Anthropic, Azure, Vertex AI, and AWS Bedrock
- Custom hosting: Build your own HTTP/WebSocket layer and run the SDK in containers
- Hooks: Easily intercept and control agent behavior
- Standardize on OpenAI: GPT-5.3-Codex and OpenAI-specific tooling
- OS-level sandbox modes: Use built-in
read-only,workspace-write, ordanger-full-accessmodes - MCP server mode: Expose your agent as an MCP server with
codex mcp-server - Observability: Use OpenAI Traces
Feature comparison
| Feature | Deep Agents | Claude Agent SDK | Codex SDK |
|---|---|---|---|
| Core tools | |||
| File Read/Write/Edit | ✅ read_file, write_file, edit_file | ✅ Read, write, edit | ✅ Read, write, edit |
| Shell Execution | ✅ execute | ✅ bash | ✅ exec |
| Glob/Grep | ✅ glob, grep | ✅ glob, grep | ✅ Built-in |
| Web Search | ✅ support for third-party and provider-native | ✅ WebSearch, WebFetch | ✅ web_search |
| Planning/Todos | ✅ write_todos | ✅ Todo lists | ✅ Plan before changes |
| Subagents | ✅ Subagents | ✅ Subagents | ✅ Multi-agent workflows (experimental) via MCP + Agents SDK |
| MCP Client | ✅ | ✅ MCP | ✅ MCP client |
| Human-in-the-Loop | ✅ Approve/edit/reject | ✅ Permission modes | ✅ Approval modes (Full Access, Read-only, Auto) |
| Skills System | ✅ Skills | ✅ Skills | ✅ Skills |
| Additional instructions | ✅ Memory | ✅ CLAUDE.md files | ✅ AGENTS.md |
| Long-term memory | ✅ Memory Store | ❌ | ❌ |
| Streaming | ✅ Streaming | ✅ Streaming | ✅ Streaming |
| Sandbox Integrations | |||
| Agent runs in sandbox | ✅ Can run in sandbox | ✅ Can run in sandbox | ✅ Can run in sandbox |
| Agent runs operations in sandboxes | ✅ Uses sandboxes as tools | ❌ | ❌ |
| OS-level sandbox modes | ❌ | ❌ | ✅ read-only, workspace-write, and danger-full-access |
| Architecture | |||
| Composable Middleware | ✅ | ✅ Hooks | ❌ |
| Virtual Filesystems | ✅ Virtual filesystem with pluggable backends | ❌ | ❌ |
| Deployment | |||
| Production hosting | ✅ LangGraph Platform, self-hosted, or standalone | Build your own HTTP/WebSocket layer; run SDK in containers | ❌ |
| Local execution environment | ✅ | ✅ | ✅ |
| Cloud execution environment | ❌ | ❌ | ✅ |
| Supported protocols | |||
| ACP server (IDE integration) | ✅ ACP server (python only) | ✅ (third-party) | ✅ (third-party) |
| MCP server mode | MCP endpoint through Agent Server | ❌ | ✅ codex mcp-server |
| A2A endpoint | ✅ A2A endpoint in Agent Server | ✅ | ❌ |
| State management | |||
| Session Resume | ✅ | ✅ Session management | ✅ Resume threads; SDK resumeThread() |
| File Checkpointing | ✅ Backends and checkpoints | ✅ File checkpointing | ✅ Git checkpoints |
| Time Travel (state branching) | ✅ | ✅ | ✅ |
| Observability | |||
| Native Tracing | ✅ LangSmith | ❌ | ✅ OpenAI Traces |
| Evaluations | ✅ LangSmith | ❌ | Evals patterns (Command-line with codex exec --json) |
Notice a mistake?
We drafted this comparison on March 4th, 2026. If you notice mistakes or changes in products, please file an issue.Connect these docs to Claude, VSCode, and more via MCP for real-time answers.