Questions we get asked.
Short answers, in plain English. If yours is not here, email hello@laminagrid.com.
What it is
What does Lamina Grid do?
We build AI agents that run business operations. They work on infrastructure you own, under rules you write, and every action they take is recorded against a person. We run our own business on the same stack.
What is an agent harness?
A harness is the runtime an AI agent runs inside: it supplies the model connection, the tools, the memory and the permissions, and it decides what the agent can reach. Ours ships each agent as a versioned bundle, so the same agent runs the same way on every machine. See how it works →
What is LAW?
LAW is a checkpoint on the model data plane. Every model call is inspected on its way out and on its way back by a small, purpose-built judge running on the same machine, so inspection adds no round-trip to another provider. LAW is in progress. More on LAW →
Do we have to change our tools or our model provider?
No. Agents keep the provider key you already use, and the harness sits around the work rather than around your stack. LAW is adopted by pointing agents at it; nothing inside the agent changes.
How is this different from a copilot?
A copilot sits beside the work and a person still does it. These agents do the work: they hold the queue, run on your infrastructure, and their actions are attributed and auditable.
Control and data
How is what an agent does made auditable?
Every action, approval and answer is bound to a validated principal and written to an append-only, hash-chained record the agent cannot reach. For any action you can answer who asked, who approved and who answered, recorded as it happened rather than reconstructed afterwards. How we govern agents →
Can an agent do something we did not allow?
No. An agent's authority is declared when it is built, not when it is prompted, so a clever instruction cannot widen it. It cannot reach a tool it does not carry, and destructive actions require a person.
Who can reach our agents?
One door. A sidecar process owns the public port, refuses anything without a valid credential and proxies only authenticated traffic inward. The agent's own port is never published.
What happens when an agent gets something wrong?
It is on the record, with the person who approved the action. Borderline and unusual cases are handed back to a person as a list, and a refused or failed check is logged rather than swallowed.
Who it is for
Which industries do you work with?
Marketplaces, finance, property, healthcare, e-commerce and professional services, mostly across Asia Pacific. Where AI pays off →