Mapier Labs — frontier interaction and agent research

Project 1 of 6: Presence

  1. Presence: An agent that shows its work while it thinks, not after.
  2. Orbit: Plans that assemble themselves from the messages you already sent.
  3. Harness: A rig for watching an agent fail slowly, across days rather than turns.
  4. Relay: One number, every channel, and the agent never learns the difference.
  5. Recall: Context an agent can defend, not a transcript it merely has.
  6. SDK: The parts we would rather not have you rebuild.
Scroll to thesis
mapier labs
DemosResearch

The thesis

Buildforagents.Measurethefailures.Shipwhatholds.

Interfaces

Agents need surfaces built for them, not borrowed from us.

A chat box is a text field with a good marketing story. We build the interactions that assume an agent on the other side — plans you can correct mid-flight, state you can see, and handoffs that do not lose the thread.

Evaluation

A benchmark is only useful if the current model fails it.

Most agent evaluations end when the task does. We measure the part that breaks in production instead: long horizons, real interruptions, and what an agent still holds after a gap of days.

Shipping

Research that does not reach a product is a draft.

Everything here has somewhere to land. What holds up moves into mapier.ai and Mapi, where real people use it and the failure modes stop being hypothetical.