Why Open Cradle
Every company is building agents. In a few years there will be millions of them. The open question is not how to build one — it is who runs them, under what rules, and who answers for what they do.
Almost every company is now building its own AI agents. In a few years there will be millions of them.
That raises a question the industry has barely started on. Not how do you build an agent — that problem is being solved, repeatedly, by dozens of teams, and it gets easier every quarter. The question is: who runs them?
Not individual models. Not individual workflows. The whole system of interaction between agents, knowledge, security policies, corporate processes and the people who remain accountable for the outcome.
The pattern is familiar
We have watched this shape before.
Applications needed a place to run, with isolation, scheduling and a stable interface to hardware. That became the operating system. Containers needed a place to run, with scheduling, identity, networking and failure handling across many machines. That became Kubernetes.
In both cases the runtime was not the interesting part when the technology was young. Early on, the exciting work was in the applications, and the runtime looked like plumbing. It became the centre of gravity later, when the number of things being run crossed a threshold where hand-management stopped working.
Agents are approaching that threshold. Ten agents can be governed with a spreadsheet and a careful team. A thousand agents — written by different teams, some vendor-supplied, some assembled by a business unit without engineering review, calling each other's tools, touching overlapping data under different legal obligations — cannot. It needs a layer where identity, permission, verification, escalation and audit exist once, for all of them.
What Open Cradle is not
It is not another AI framework. Frameworks help you write an agent; the hard part starts after the agent works.
It is not an agent builder. Composing steps into a workflow is the visible part of the problem and the least durable.
It is not an AI Factory. Documents in, index, model, assistant out — the first post in this series is about why that pipeline stops at the demo.
It is a harness system for intelligent systems: a runtime where agents propose, the domain model validates, policy decides what is permitted, tools execute, verification confirms, and a human owns the decisions that need an owner.
What this changes in practice
For an organisation, the shift is from "we deployed an AI assistant" to something with a different shape:
You can state what an agent is allowed to do, and enforce it somewhere other than a prompt. You can answer why a decision was made with a chain of evidence rather than a plausible paragraph. You can put the human at the points that carry consequence, instead of either reviewing everything or reviewing nothing. You can add a second agent, and a tenth, without the governance cost growing with each one. And you can do it on your own infrastructure, which for many of our users is not a preference but a legal requirement.
That last point deserves its place, and not the headline. Self-hosting matters — the whole platform runs on your hardware, in your network, under your control, and can be air-gapped. But running locally does not by itself make a system safe. A hallucination on your own server is still a hallucination. Sovereignty is a property of where the system runs; the harness is a property of how it runs. You want both, and only one of them is commonly on offer.
Where we are, plainly
Cradle today does the execution part: local model orchestration, agent routing, grounding in your own documents, a two-tier risk gate, human approval before anything leaves the perimeter. That works, on-premise, now.
The rest of what this series describes — the policy engine, the verification runtime, the ontologies and domain packs as first-class components — is the architecture we are building towards, documented publicly and marked as such in the architecture reference. We would rather be measured against a design we have committed to in writing than sell a finished system that does not exist yet.
The bet
Our bet is that within a few years, the interesting question about enterprise AI will not be which model a company uses. Models will be interchangeable, and mostly good enough.
The interesting question will be what runs between the model and reality — and whether an organisation can explain, bound and stand behind what its agents do.
That layer needs to exist. We think it should be open, self-hostable, and designed for the organisations that carry real accountability rather than for demos.
That is why Open Cradle.
This was the final post in a five-part series: AI Factories · RAG · Harness systems · The neuro-symbolic runtime · Why Open Cradle.