Fleet RLM overview

• DSPy runs each Turn through a direct, flex, or rlm route.
• Daytona sandboxes run every line of model-authored Python, outside the Fleet process.
• PostgreSQL stores durable Session state: Turn history, file content, memories, and Session selections.
• MLflow records traces and runs evaluation. Fleet uses a local MLflow server by default.

​

What you can do

• See the work as it happens. The terminal shows routing, generated Python, tool activity, child sandboxes, and the final answer in one timeline.
• Continue a Session. History, workspace files, reviewed memory, and model and mode selections survive a restart. Python variables in the sandbox are not durable.
• Work with large files. Large inputs stay in workspace files. The model receives paths, metadata, and bounded previews, then explores the data through Python.
• Delegate independent work. spawn_agent runs a subtask in a fresh child sandbox. spawn_agents runs independent subtasks concurrently within shared limits.
• Evaluate and optimize. Label traces, score Turns, and compile GEPA candidates with fleet eval and fleet optimize.
• Drive it from your own client. The same server runs without the terminal and exposes an HTTP and SSE API.

​

How the pieces fit

Terminal client ── HTTP / SSE ──> FastAPI ──> DSPy Turn program
                                    │              │
                                    │              └── Daytona: generated Python,
                                    │                  root and child sandboxes
                                    ├── PostgreSQL: durable Session state
                                    └── MLflow: traces and evaluation

​

Next steps