Two products. One platform.

A runtime that ships agents as pinned bundles, with the fleet Portal built into it, and a firewall on the model data plane.

Product 01 · Shipped

Lamina Grid Harness

Agents you ship like software. Versioned, reproducible, and operable from one window, not servers you hand-configure and hope stay the same.

Ship, don't babysit

One pinned bundle per role (base, runtime, plugins and digest), so every agent is reproducible and nothing drifts between them.

The fleet in one pane

The Portal is built in: every agent, its skills, its memory and the actions you allow, without a second console to deploy.

Work that runs itself

Recurring jobs live in versioned Markdown orders that fire on schedule and read like documentation: reviewable line by line.

01

Run the bundle

Pick the bundle for the role and start it. Nothing installs into the host.

02

Operate from the Portal

Find the agent, read its memory, then chat, restart, or hand it work.

03

Schedule the rest

Write a standing order. It arms its own wake and runs.

The Lamina Grid Harness running a live session: sidebar, an agent conversation with tool calls, and a files panel

Working with an agent: one session, its files, and every tool call it made.

The Grid page inside the harness: Organisation Dashboard with 17 agents across 12 sub-groups and 7 groups, agent names blurred

Operating the fleet: every agent grouped by team, with live status and one click to any of them.

Product 02 · In progress

LAW (Agent Wall)

Every call reviewed. Nobody notices. Inspection inside the request path, at a speed the user never feels, with a record you can hand to an auditor.

Big-model judgement, small-model speed

A customised judge tuned to a far larger model's call: frontier-quality decisions at a fraction of the compute.

No second round-trip

The judge answers on the same box in one forward pass, and no database sits in front of the decision. Latency stays flat as the fleet grows.

No keys, no blind spots

It never holds your credentials, the agent's own key passes through, and every decision lands in a hash-chained audit.

01

Point the agents at it

Change the provider URL to the appliance. Nothing inside the agent changes.

02

Write the policy

Per-agent files; a revoked rule applies without a restart.

03

Read the record

Decisions in the audit view; a refusal reaches the agent with a reason it can read.

Agentsany harness
→
LAWpolicy · checks · judge
→
Provideragent's own key
The judge answers on the same box in one forward pass: loopback, not a second provider call
Audit is written after the decision, so it never sits in front of the reply

Nothing on the decision path leaves the appliance. Rust core, stateless per request; the reply leg records verbatim and can cut the stream.

What ships inside the harness

The three you will actually use. The rest is plumbing.

Dreammaker

Persistent memory. Sessions are harvested into a knowledge base, and the index goes in front of the agent every turn.

Standing Orders

Recurring work written in versioned Markdown. It arms its own wake and runs on schedule.

The Grid

The fleet directory inside the product: every agent and the capabilities it declares.

Run the platform yourself

We use both to run our own operations, including the property platform on top of them.