Can Deep Agents Resume Work After Sandbox Loss?

Run a two-stage LangChain Deep Agents job through your cloud backend and measure recovery from workspace loss. Use LangGraph tracing to locate replay and NVIDIA NemoClaw as a separate trial if you want to host the agent inside the sandbox.

Choose where the agent loop should run

Start with the placement decision. An application-owned harness can keep model calls and orchestration in your service while forwarding workspace operations to a cloud execution backend. If you want to evaluate hosting Deep Agents Code inside the sandbox, NVIDIA NemoClaw documents an interactive installer path that lets you choose the agent and inference provider. Use the same tiny fixture for both layouts. Ask the agent to transform a synthetic CSV and label three events in the operator trace: model response received, sandbox command finished, and output returned to the caller. Record which component emits each event and where the CSV was written. If the command completes but the caller receives no output, those observations tell you which boundary to investigate before you retry the whole agent run.

GitHub - NVIDIA/NemoClaw: Run agents like Hermes, LangChain Deep Agents, and OpenClaw more securely inside NVIDIA OpenShell with managed inference · GitHub

Measure replay at a deliberate checkpoint

LangGraph provides durable execution for the orchestration layer underlying Deep Agents. Test workspace recovery separately with a two-stage CSV job. Have stage one write a deterministic intermediate file and record its digest in the harness checkpoint. Have stage two read that file and produce a sorted report with a known digest. At the stage boundary, run one control to completion and run a second trial that resumes from the saved checkpoint against a fresh sandbox. Count stage-one executions and measure from the resume request to a verified report. A missing intermediate identifies a workspace dependency. A matching report with two stage-one executions means the run recovered through replay. If replay exceeds your deadline, explicitly restore the intermediate before resuming and rerun the same fixture.

GitHub - langchain-ai/langgraph: Build resilient agents. · GitHub

Sources