Deep Agents Deploy: an open alternative to Claude Managed Agents

The agent platform wars have a new front, and it's not about which LLM is smartest. Deep Agents Deploy and Claude Managed Agents share nearly identical architectures—orchestration layer, sandbox environments, multi-protocol endpoints—but diverge sharply on memory ownership. For production teams, this distinction matters more than model performance benchmarks suggest.

Both platforms solve the operational complexity of taking an agent from prototype to production. You specify your model, provide instructions via AGENTS.md format, attach skills (specialized knowledge as markdown files plus executable scripts), configure MCP tools, and select a sandbox provider. A single deploy command spins up a horizontally scalable server with 30+ endpoints supporting MCP, A2A, and Agent Protocol. The deployment story is table stakes at this point.

The critical difference is what happens to agent memory over time. Deep Agents stores memory in standard formats—AGENTS.md files, skill definitions, conversation context—that you can query directly via API or self-host entirely. Claude Managed Agents keeps memory behind Anthropic's API. This isn't an abstract philosophical concern about openness; it's a concrete operational risk.

Consider the switching cost calculation. Moving between LLM providers is genuinely not that hard anymore. You adjust some prompts, maybe tune temperature settings, run your eval suite, and you're done. Teams migrated from GPT-4 to Claude 3.5 Sonnet en masse when the benchmarks justified it. Model APIs alone don't create meaningful lock-in because the interface is commoditized and the context window is stateless per request.

Agent memory is different. An internal SDR agent that learns customer preferences, discovers which objection handling patterns work, and accumulates institutional knowledge about your sales process—that memory is the actual product. If it's locked behind a vendor API, switching harnesses means resetting to zero. For customer-facing agents, the problem compounds. A support agent that remembers customer history, preferences, and past issue resolution creates a data flywheel that improves experience over time. If that memory lives in Anthropic's infrastructure, you don't own the flywheel.

The practical question for platform teams is whether Deep Agents' openness creates operational burden that offsets the portability benefit. The model-agnostic design means you're responsible for managing provider credentials, handling rate limits across different APIs, and dealing with inconsistent response formats. Claude Managed Agents abstracts all of that away, but you're betting that Anthropic's model lead persists indefinitely and that their pricing remains competitive.

Sandbox integration is another tradeoff surface. Deep Agents supports Daytona, Runloop, Modal, and LangSmith Sandboxes out of the box, with an implementation guide for custom providers. This flexibility is valuable if you already have containerized execution infrastructure, but it means you're stitching together multiple vendor relationships. Claude's integrated sandbox is simpler operationally but gives you no leverage if their sandbox pricing or performance becomes a bottleneck.

The self-hosting option for LangSmith Deployments addresses compliance and data residency requirements that make managed services non-starters for some enterprises. If your agent processes PII or proprietary business logic, keeping memory in your own databases isn't optional. The operational cost is real—you're running and monitoring the deployment infrastructure—but the alternative is shipping sensitive data to Anthropic's servers.

For teams building internal tooling with limited external exposure, Claude Managed Agents' convenience may outweigh lock-in risk. For customer-facing agents or scenarios where memory constitutes competitive advantage, Deep Agents' architecture makes the portability-complexity tradeoff worth it. The memory ownership question isn't theoretical; it determines whether your agent's accumulated knowledge is an asset you control or a liability that binds you to a single vendor.