In April 2026, two major players dropped significant agent frameworks within 24 hours of each other:
Both solve the same core problem: making AI agents that can actually do things — plan, execute, remember, and adapt across long, multi-step tasks. But they take fundamentally different approaches.
One gives you control. The other gives you convenience. Choosing the wrong one could cost you weeks of engineering time.
Basic LLM tool-calling agents are shallow. They work for simple tasks but fall apart on complex, multi-step work:
Both frameworks were built to fix these problems. They just do it differently.
LangChain Deep Agents is an open-source agent harness built on LangGraph. It packages four key primitives for long-horizon, non-deterministic tasks:
1. Planning (write_todos) — decomposes complex tasks into trackable steps, adapts as work progresses
2. Subagent Spawning (task tool) — delegates subtasks to isolated subagents, each with their own context window
3. Virtual Filesystem — stores large artifacts outside the context window, available across the full task lifecycle
4. Memory & Persistence — LangGraph's Memory Store provides cross-session recall
| Feature | Details |
|---|---|
| Model agnostic | OpenAI, Anthropic, Google Gemini, Ollama, and more |
| Human-in-the-loop | Built-in pause/resume with human input |
| Streaming | Real-time output via LangGraph |
| Observability | Full tracing via LangSmith |
| Open source | MIT license, full code control |
On April 9, 2026 — one day after Anthropic launched Claude Managed Agents — LangChain launched Deep Agents Deploy in public beta, explicitly positioning it as an open alternative.
AGENTS.md ← Agent definition (open standard)
/skills ← Skill definitions (open standard)
mcp.json ← MCP server config (convention)
This mirrors the configuration format used by Claude Code and Cursor — immediately familiar to developers in the agentic coding space.
deepagents deploy
The CLI packages your agent and deploys it to LangSmith Deployments, automatically provisioning assistants, threads, runs, memory stores, and checkpointers.
Models: OpenAI (GPT-4o, o3), Anthropic (Claude Sonnet, Opus), Google (Gemini 2.0, 2.5), Azure OpenAI, AWS Bedrock, Fireworks, Ollama
Sandboxes: Daytona, Runloop, Modal, LangSmith
Every deployed agent is automatically exposed via three open protocols:
| Protocol | Purpose |
|---|---|
| MCP (Model Context Protocol) | Expose agent as a tool callable by other agents or clients |
| A2A (Agent-to-Agent) | Enable multi-agent setups where agents communicate directly |
| Agent Protocol | Standard interface for agent interoperability across frameworks |
No vendor lock-in. Any MCP-compatible client, A2A orchestrator, or Agent Protocol-compliant system can call your agent.
from deepagents import create_deep_agent
from deepagents.backends import CompositeBackend, StateBackend, StoreBackend
agent = create_deep_agent(
backend=CompositeBackend(
default=StateBackend(),
routes={"/memories/": StoreBackend(...)},
),
system_prompt="You have persistent memory at /memories/..."
)
Your data lives in your databases — not Anthropic's servers. Critical for enterprises with data residency requirements.
Anthropic's fully hosted agent service, launched April 8, 2026. The philosophy: decouple the agent's brain (Claude model) from its hands (execution environments). Anthropic manages the runtime, sandboxes, session persistence, and security. You define what the agent does.
Agent — reusable config: Claude model, system prompt, tools, MCP servers, skills
Environment — cloud container: pre-installed packages, network rules, mounted files, isolated sandboxes
Session — persistent instance: append-only event log, server-side history, fetch anytime
Events — SSE streaming: user inputs, tool results, status updates, model responses
| Feature | Details |
|---|---|
| Fully hosted | No infrastructure to manage |
| Secure sandboxes | Credentials isolated from untrusted code |
| Built-in harness | Agent loop, tool execution, prompt caching, compaction |
| Async operation | Long-running tasks without keeping connections open |
| MCP support | Native Model Context Protocol integration |
| Sub-agents | Spawn specialized agents for coding, research, design |
| Dimension | Deep Agents Deploy | Claude Managed Agents |
|---|---|---|
| Hosting | LangSmith Deployments | Anthropic Cloud |
| Model support | Any (OpenAI, Claude, Gemini, Ollama) | Claude only |
| Configuration | AGENTS.md, /skills, mcp.json | Agent + Environment + Session API |
| Protocols | MCP + A2A + Agent Protocol | SSE events only |
| Memory ownership | You own it (your databases) | Anthropic-managed |
| Sandbox providers | Daytona, Runloop, Modal, LangSmith | Anthropic cloud containers |
| Open source | Yes (MIT license) | No |
| Pricing | LLM costs + LangSmith fees | LLM costs + $0.08/session-hour |
| Vendor lock-in | None | High |
| Setup | One CLI command | API-based setup |
| Data residency | Your infrastructure | Anthropic's infrastructure |
| Security sandboxing | Provider-dependent | Built-in, Anthropic-managed |
| Observability | LangSmith (excellent) | SSE event logs |
LangChain's timing was deliberate. By launching Deep Agents Deploy one day after Claude Managed Agents, they sent a clear message:
"You don't have to choose between managed infrastructure and open standards. You can have both."
Anthropic's bet: The future of agents is safe, managed, and Claude-powered. Security and reliability handled at the infrastructure level.
LangChain's bet: The future of agents is open, composable, and model-agnostic. Infrastructure that any team can customize, extend, and own.
Both bets could be right. The market is large enough for both approaches.
✅ You need model flexibility — GPT-4o today, Gemini tomorrow ✅ Data ownership is a requirement (compliance, residency) ✅ You want open standards — MCP, A2A, Agent Protocol ✅ You're building internal tools where open-source is required ✅ Vendor lock-in is a concern ✅ You have or are building LangGraph expertise
✅ You're already Claude-first and want to go deeper ✅ You want to ship in days, not weeks ✅ Security and sandboxing are critical (legal, finance, healthcare) ✅ You're building production agents without DevOps overhead ✅ Your tasks are knowledge work — research, analysis, document processing ✅ Async, long-running agents are your primary use case
Many teams will use both:
| If you value... | Go with... |
|---|---|
| Flexibility & model choice | Deep Agents Deploy |
| Speed to production (Claude-first) | Claude Managed Agents |
| Open standards (MCP, A2A) | Deep Agents Deploy |
| Security & sandboxing | Claude Managed Agents |
| Data ownership | Deep Agents Deploy |
| Managed infrastructure | Both (now comparable) |
| Open source | Deep Agents Deploy |
| Simplest setup for Claude users | Claude Managed Agents |
The agent infrastructure war is officially on. Both platforms launched within 24 hours of each other, both targeting the same production agent use cases, both with managed deployment. The key differentiator is now openness vs. polish — and that's a choice only your team can make.
Sources: LangChain Deep Agents Deploy launch blog (blog.langchain.com, April 9, 2026), LangChain Deep Agents docs (docs.langchain.com/oss/python/deepagents/deploy), Anthropic Claude Managed Agents launch blog (April 8, 2026), LangSmith observability docs, CRAB benchmark leaderboard, Anthropic customer case studies (Fountain, CRED, Legora), Deep Agents v0.5 release notes.
