Self-tuning agent harness on MongoDB Atlas

Give every user an agent that fits them.

Molt stores every run as a trace. An evolver turns failures into a small, cited patch to the harness's rules, context policy, guardrails and tool access, replays the user's evals on the candidate, and only a human promotes it. The result is a versioned, auditable harness per user.

See the hallmarks

Users

One base agent. One harness per user, versioned and eval-scored.

How it works

  1. 1. Run

    Retrieve memories under the context policy, assemble, call the model, apply guardrails. Everything lands in a trace.

  2. 2. Feedback

    The user says what was wrong. It is stored on the trace and written as an episodic memory.

  3. 3. Evolve

    The evolver turns failures into a small, cited patch across rules, context, guardrails and tools.

  4. 4. Gate

    Six deterministic evals replay on current vs candidate. Locked guardrails cannot be weakened.

  5. 5. Promote

    A human promotes all or some operations. Every version is kept and can be rolled back.

The evolver is AI. The evals and the gate are deliberately not. The harness evolves config only, never code.