DeepAgent Sandbox Codebase Dossier

DeepAgent Sandbox is an agentic data-analysis platform in a planned/partially implemented state. The documented target combines a React frontend, a product backend, LangGraph/DeepAgents orchestration, a Fastify microsandbox executor, Postgres checkpoints/metadata, MinIO artifacts, and optional Redis.

Read first

• Evidence map
• Product overview
• Agentic architecture
• Executor backend
• Risks and implementation gaps

Important qualification

The repository README describes the target architecture, while the checked-in implementation visibly contains a substantial microsandbox-executor and frontend API client. The planned Python backend/ and LangGraph graph described in implementation_plan.md are not present in the inspected tree. This dossier distinguishes implemented executor behavior from planned product behavior.

flowchart TD
  classDef default fill:#1e293b,stroke:#38bdf8,stroke-width:2px,color:#f8fafc
  classDef highlight fill:#065f46,stroke:#34d399,stroke-width:2px,color:#f0fdf4
  User["Analyst"]:::highlight --> UI["React chat/artifact UI"]
  UI --> Planned["Planned product backend"]
  Planned --> Graph["LangGraph + DeepAgents"]
  Graph --> Executor["Fastify microsandbox executor"]
  Executor --> VM["MicroVM runtime"]
  Executor --> Store["MinIO / local storage"]
  Graph --> PG["Postgres checkpoints"]
  linkStyle default stroke:#64748b,stroke-width:2px

Top facts

• The plan explicitly separates reasoning from execution control.
• The executor currently exposes sessions, jobs, runtime leases, storage sync, health, and network policy.
• JobExecutor validates limits, serializes per-session jobs, stages workspace state, executes Python/Bash, and persists diffs.
• The frontend already models auth, threads, files, messages, run events, uploads, artifacts, and SSE.
• Presigned upload/download, authenticated backend routes, and DeepAgent backend integration are target design items.
• A source-of-truth reader must not mistake the target plan for completed functionality.