Design partner program is openThree partners, real decision traffic →
razoo

Our mission

Make AI agents earn their autonomy. Then make them cheaper every week.

Every agent in production today is exactly as dependent on the model, and exactly as expensive, on day four hundred as on day one. The judgments people make around it are thrown away after each decision. Razoo exists to keep them, certify them, and let them run for nothing, under authority a person can take back in one click.

The industry is building better agents and better fences around them. Neither changes the economics. An agent asks the model the same question at the same price every day, a person confirms the same answer, and nothing is learned from the confirmation. The cost scales with usage forever, and when the agent is wrong, nobody can say who allowed it.

Cost that never flattens

Model spend is now a board-level line. It grows with every user, every workflow, and every retry, and no amount of prompt engineering changes the shape of that curve.

Judgment that is thrown away

The most valuable output of an agent in production is the stream of human decisions around it. Today that stream is discarded the moment each decision is made.

Authority nobody can account for

When an agent pays, sends, files, or signs, the question is not what the model thought. It is who allowed it, under what limit, and whether it can be undone.

Why now

Agents moved into operations. The governance layer that followed answers the wrong question.

In 2025 agents left the demo and entered accounts payable, contracts, support, and the inbox. In 2026 the control planes arrived to manage them: well funded, well built, and focused on permission. What agents may do. Which credentials. Which budgets. Necessary, and static.

Control planes ask

May this agent do that?

  • Unit: an agent, a credential, a budget.
  • Changes when a person edits policy.
  • No different after a year of running.
  • Sits above the agent.

Razoo asks

Does this decision still need a model at all?

  • Unit: one decision, with its evidence.
  • Changes with every judgment a person makes.
  • Improves with use.
  • Sits under the agent, before any side effect.

Razoo runs under a control plane, not instead of one. The gap it fills is the one every control plane leaves open by design: the economics and the learning. A firewall is a firewall. A flywheel is better every week it runs.

Spend

Agentic workloads call the model several times per task, and the bill grows with every task.

Law

We expect regulation to keep moving toward mandatory human oversight and logging for high-risk AI systems. We build to that standard now rather than wait for it.

Proof

We believe the observe, certify, promote, demote loop follows a pattern with precedent in published research. As far as we know, nobody has shipped it as a kernel under the agent.

The corridor

The rest of the stack holds agents where they are. Razoo moves them.

Control planes fence agents from above. Frameworks and models make them stronger, with the model in the loop forever. Workflow tools sit below, deterministic and audit-ready, reached only by hand-built steps. Razoo runs the corridor between the two worlds: certified decisions move down and become software, doubt moves back up to the model.

The four-question test

Four questions. Most tools answer yes to one.

Learn from what agents actually do, and turn proven behaviour into software that runs without the model. Take that authority back when the world drifts, and produce a trail an auditor can replay.

Learns from what agents actually doTurns proven behaviour into model-free softwareTakes authority back on driftProduces authority an auditor can replay
RazooYesYes, certified by replayYesYes
Control planesNoNoNoLogs only
Approval toolsNoNo, humans approve foreverNoPartial
Workflow automationHumans onlyManual buildsNoNo
Model providersYesNo, learning stays in the modelNoPartial
Tool and skill layersYes, pooled across customersNo, the model stays in the loopNoNo
Cache and replay toolsCachesReplays without verificationFallback onlyNo

categories only, no vendors named · full comparison in the technical review

What has been built

A kernel. Built, tested, and running.

Everything below is running code. Roadmap items are labelled on the pages that carry them. The claims below are enforced by the software and pinned by automated tests that fail if a guard is removed.

A resolution ladder with the model on the bottom rung

Deterministic rungs first, learned rules next, the model only for what is left, a person for anything the model may not decide. One order. No exceptions.

Authority earned by replay

A learned rule gains authority only by replaying cleanly against the real judgments people already made. Confidence scores grant nothing.

Authority bound to what it depends on

Every rule is bound to the registry entries it relies on and re-checked before any side effect. When they change, it stops firing before anyone notices it should.

A machine-checked execution rule

Whether an action may execute is decided in one place, and that decision is machine-checked every time the code changes.

Revocation as a first-class act

A rule is revoked on drift or counterexample, or by a person, and loses force everywhere at once. Work returns to a person. Nothing is dropped.

Bit-for-bit replay

No ambient clock or random source on a decision path, enforced by a lint in CI. Any decision replays against the current kernel and compares, which is how upgrades are checked before they apply.

Two deployment shapes

A single-tenant appliance on the customer's infrastructure, and an SDK that puts the kernel inside a product. Same guarantees, same code.

Observe and the crystallization report

A read-only proxy that records an agent's decisions without changing them, and a report that says which of them are repeats. The falsification instrument for the whole thesis.

GovBench

A seeded, deterministic benchmark for governed execution that any stack can run against its own adapter. Same seed, byte-identical report.
every invariant pinned by a test internal adversarial pass, findings pinned by tests determinism lint in CIWhat the technical review covers

The thesis

The moat is not the model. It is the evidence.

Models get cheaper and better every quarter, and every vendor has access to the same ones. What no vendor has is a customer's own history of decisions: which invoices this company approves, which contracts it escalates, how its people reply. Razoo turns that history into certified, revocable authority, and the longer it runs the more of a customer's work it carries for nothing.

That is a compounding asset that lives with the customer and cannot be copied by a competitor with a better prompt. It is also why the design partner program measures one thing first: the share of governed decisions that stop needing the model. The number is the pitch. The architecture is how it is reached.

The commercial shape follows from it. Razoo is packaged software under a commercial licence: the appliance image, updates and support by scope, and an embed agreement that covers every tenant a vendor has. No tier includes source. Priced on the decisions the kernel counts. No seat pricing, and no fee that shrinks when the product works.

What we believe

Autonomy should be earned, bounded, and revocable.

Evidence over confidence

A model that is ninety-nine percent sure is still a model. A rule runs on its own only after it has replayed cleanly against the judgments people already made.

Every rule fires inside an envelope

The envelope names the fields, the actions, and the limits a rule may touch. It can narrow what an action allows and never widen it. Money and legal fields always need a person.

Taking authority back is one act

Revoked on drift or counterexample, or by a person, and the rule loses force everywhere at once. Nothing is dropped and nothing is silently restored.

The company

Small by design. Held to the same rules as the product.

Razoo is built by Oryn Technologies Pty Ltd in New South Wales, Australia. The team is deliberately small and works with AI coding agents under the discipline the product sells, because we could not credibly sell it otherwise.

A written constitution

One document every contributor and every coding agent reads first. It names what may be touched, what may never be done, and the design law that every process step is plumbing, reasoning, or judgment.

A definition of done

Typecheck, the determinism lint, and the full test suite green, plus the specific review items for that change. A task without a done condition is not handed over.

Independent verification of every commit

One prompt, one session, one reviewable commit, verified against its review items before the next is issued. Nothing lands on the word of the agent that wrote it.

Measured numbers only

No customer, quote, logo, certification, or benchmark result appears anywhere unless it exists. When the conversion number exists, it will be published with its evidence.

Design partners

Three slots, one per workflow, with real decision traffic. Free during the program and preferred terms in writing at general availability.

Advisors and investors

We are assembling an advisory group across enterprise operations, AI infrastructure, and Australian and US go-to-market, and speaking with investors who back infrastructure companies at the seed stage.

Intellectual property

A provisional patent application covering the certification and revocation mechanisms is being prepared. The code is proprietary and never distributed: customers receive packaged images and tarballs under a commercial agreement, and audit it through the technical review under NDA.

The name

Not a brass razoo.

In Australian slang a razoo is a coin of no value at all. Nobody has ever held one. It is what a decision costs once an agent has earned the right to make it without the model. Our mark is that coin, with the three rungs of the ladder cut into it: the model at the top, learned authority in the middle, and the deterministic rule at the bottom, wide and brass.

Design partners bring a workflow. Investors bring a question.

Either way, the conversation starts with the mechanism and ends with the number.