The governed record for the autonomous company · BYOC

The management layer
was never a system.

Every company runs software for its functions. None of it holds the layer above — the decisions and why they were made, the obligations being carried, who approved what and on whose authority. lagstyr makes that layer a system of record, and makes it the only surface an agent can act through.

Request early access How it works

The gap

Your systems record what happened. Nothing records why.

the gap · 01

The unwritten layer

Decisions, obligations, approvals, risks, the evidence behind a claim — none of it has a home. It lives in email threads, board packs, spreadsheets, and the heads of the four people who have been here longest.

the gap · 02

And now agents are acting

The layer carried by human judgement is exactly the layer agents need in order to act — and exactly the layer that must be auditable once they do. A record that was good enough for people is not good enough for machines.

lagstyr makes that layer a system of record — one identity, one authority, one memory — and the single governed surface anything, human or agent, must pass through to change it.

The brain of the organisation

Not a database. Not an automation layer.
The organisation's brain.

Everything the organisation knows lives in one place and compounds. Knowledge is built, not filed. Obligations are tracked, not remembered. Risk is watched continuously, not audited once a year.

  1. one identity

    Every entity has a single canonical id — across every system it appears in. Everything links back to it, and that linking is the universe.

  2. one authority

    Every fact has one authoritative home and names the source it came from. Authority is sacred; projections are disposable and rebuildable at will.

  3. one memory

    An append-only log records what changed, when, and on whose authority. No row is ever written without naming its source.

  4. one action surface

    Every agent action passes a single governed, typed chokepoint. Agents are never pointed at raw tables.

An operating system does not do the applications' work. It decides what may run, with what authority — and keeps the record.

What it owns

The records no other system holds.

lagstyr is authoritative for the management layer — the objects that have never had a home in software, and the ones an auditor, a board, or a regulator asks for first.

  • Decisions
  • Obligations
  • Approvals
  • Risks
  • Controls
  • Incidents
  • Evidence
  • Assertions
  • Board & resolutions
  • Cap table & equity
  • Agent actions
  • Measurement

Everything else — the ledger, the payments rail, the mailbox, the repository — stays exactly where it is.

Wrapped, not replaced

Your systems stay where they are.

lagstyr does not replace your ledger or your CRM, and it does not ask you to migrate. It reads your systems as sources, and acts back into them through governed seams.

inbound

Sensors

Every fact lagstyr holds names the system it came from and how far that source can be trusted. Your accounting ledger, your payments rail, your mailbox, your repository — read with provenance, never silently copied.

outbound

Actuators

When a change is approved, lagstyr acts through the same systems you already run — idempotent, routed by risk tier, and written to the record as an event before anything leaves the boundary.

The record above them is lagstyr's — and it is the one an auditor, a board, or a regulator can actually read.

Deterministic execution, driven by agents

Agents execute. Humans decide.

Deterministic substrate, probabilistic agents. They reason over projections and evidence, but their output becomes authoritative only by passing a governed write surface — so the organisation's truth is never probabilistic, and nothing material happens without a human.

propose

An agent drafts a change. It cannot commit it.

approve

A human reviews and authorises — re-authenticated, proposer excluded.

commit

Only then does the kernel write it, and emit the event in the same breath.

read

Look, never touch.

write_safe

Routine writes, deterministically spot-checked into a review queue.

write_governed

Proposed by an agent, then approved by a human.

write_kernel

The organisation's spine. Approval required; the proposer can't sign off.

An agent starts with nothing — no shell, no filesystem, no ambient reach. It is handed exactly the skills its role requires. The confine is the grant.

Full sovereignty · BYOC + BYOM

Your organisation. Your cloud. Your model.

Each deployment runs in your own cloud, on the models you choose. The intelligence and the record both stay inside your boundary — nothing leaves, and no vendor reaches in.

BYOC

Bring your own cloud

Kernel, workers, database, and storage run inside your own cloud account. Your record data never crosses its trust boundary — isolation by deployment, not by row.

BYOM

Bring your own model

The intelligence runs on the models you choose — your endpoints and keys, self-hosted or a provider you trust. Every agent run pins its model at admission; no prompt or record leaks to a vendor default.

No reach-in

No vendor on the runtime path

No licence check, heartbeat, or telemetry on any runtime path. The data plane runs fully with the vendor unreachable — your organisation doesn't depend on ours being online.

Most AI tooling gets thinner as the models improve. This gets thicker. A better model is trusted with more consequential work — which is exactly when a deterministic kernel and one typed chokepoint are worth more, not less.