Autonomous agents that do the work, and prove it.

Lunar is an applied AI lab engineering the next generation of human-machine coordination.

An applied AI lab in San Francisco.

Physical work is the next thing to automate. And autonomy is a trust problem before it is a capability problem.

So we build agents that prove it. They act, then confirm the outcome against evidence before they claim it. Our research makes that trust earnable and measurable.

Our interests span

#01 Align

Model
alignment

An agent that can act is only safe if it wants what we want. We make that true before it acts alone.

#02 Exploit

Reward
hacking

Tell an agent what to optimize and it finds the shortcut you never meant. We study the gap between the metric and the goal.

#03 Reason

Faithful
reasoning

The reasons an agent gives should be the reasons it actually used. A model can be right for reasons it will not show you.

#04 Calibrate

Calibrated
confidence

An agent should know what it doesn't know, and say so. Confidence without grounds is how a quiet error becomes a decision.

#05 Context

Context
engineering

What an agent remembers, forgets, and is allowed to see decides what it does next.

#06 Verify

Grounded
verification

Before it reports that real-world work is done, the agent proves it against evidence. A claim is not completion.

Morsa

An autonomous chief of staff for manufacturing.

The first place the work starts running itself.

Morsa coordinates people, vendors, and machines over the channels they already use, turning scattered conversation into one live view of what's done, what's blocked, and who owes what.

Backed by

Autonomous agents that do the work, and prove it.

Our interests span

Modelalignment

Rewardhacking

Faithfulreasoning

Calibratedconfidence

Contextengineering

Groundedverification