Safe-harbor infrastructure for autonomous agents

Bring your agents home safe.

Autonomous agents are starting to take consequential action on their own — unattended, at the edge, out of reach. Lightshore is the layer underneath them: it records what they did, recovers them when they break, and proves you did your diligence.

Local-firstFail-closedTamper-evidentNo cloud required
The thesis

When an agent acts alone, three questions decide everything.

Every autonomous agent that takes action unattended raises the same three questions the moment something goes wrong. Lightshore builds the primitives that answer them — each one small, deterministic, and independent of the agent it watches.

01

What did it do?

02

Can you bring it back?

03

Can you prove you were diligent?

Ship's manifest

The safe-harbor suite

What did it do?Tamper-evident local record + safety layer for every decision.
Available
Lightship
It broke in the field — bring it back to health.Deterministic on-device recovery supervisor.
In development
Keel
Was it the certified, untampered agent?Attestation of model, policy, and code integrity.
Charted
Clutch
Could you have stopped the catastrophic action?A deterministic, fail-closed brake at the action boundary.
Charted
Charter
Was it authorized to do that?A signed, machine-readable grant of authority and bounds.
Charted
Underwriter
Can you prove diligence to a court, regulator, or insurer?Turns the record into insurability — the real safe harbor.
North star
Available now

Ballast — the record that holds up.

Records every decision, flags dangerous intents, and keeps a hash-chained, tamper-evident trail — so you can prove what happened, not just hope. No cloud. Runs anywhere Python does.

Get Ballast on GitHub
One record from the trail "kind": "tool_call", "decision": "BLOCK", "reason": "rm -rf on protected path", "prev": "a1e6...42d0", "hash": "9ba2...7c11"