Preview
Learn more
Agentic Fleet
programmed to adapt, built to remember.
Agentic AI capabilities that execute, emerge, evolve, and organize themselves around how your team actually works.
Less looping.
More compounding progress.
Where RLM and stateful system meet.
A system that can keep state and build on it.
Recursive reasoning with long memory.
The system separates conversational reasoning from deep code execution. A `dspy.ReAct` Supervisor handles user interaction, planning, and tool selection. When engineering work is needed, it delegates to an `RLMEngine` that recursively writes, executes, and self-corrects Python code inside an isolated Modal sandbox.
Unlike ephemeral sandbox environments (Deno/WASM, Docker), fleet-rlm uses Modal Volumes that persist across sessions. Files written in iteration N=1 survive to iteration N=100 and beyond server restarts. Combined with Neon pgvector for semantic memory, the system accumulates knowledge over time.
