Agents are going to run the work. We make them answerable.
Within a few years the operational work of every serious company runs through agents: the claims queue, the refund line, the overnight reconciliation, the phone that rings at 9am. Whoever runs them will be asked what they did, who authorisedauthorized it, and whether they still work. AltairaLabs builds Omnia, the platform that lets you answer.
StandardiseStandardize agent behaviourbehavior, not infrastructure.
- 01
The behaviourbehavior isn’t in the code.
An agent is its prompts, its tools, its policies, its memory and its tests. That’s an artefactartifact, and an artefactartifact deserves a version number, a diff and a rollback.
- 02
The runtime is interchangeable. The record isn’t.
Your cluster, AgentCore, Vertex Agents, Foundry: one definition, four targets. What you keep is the ability to say what ran, under which rules, and run it again.
- 03
Testing that stops at go-live is theatretheater.
Models change underneath you. Suppliers deprecate. The cases that passed on Friday are run again on Tuesday, against every model you allow, on a schedule you set.
Every agent in the company on one plane, whichever cloud each one landed on. Ten thousand conversations a night, each one simulated from the customer’s side, scored, and kept if it passed. An auditor asks who authorisedauthorized one transaction and the answer is a record with a policy version on it.
The model is the easy part. The hard part is everything an enterprise has always needed from software it trusts with money: a definition it can read, a version it can name, a rule it can point to and a test it can rerun. Omnia is that, for agents.
Charlie Holland
Founder
Thirty years as a cloud and platform architect, most of it inside regulated estates: global banks, top-five pharma, heavy industry. Kubernetes and all three clouds, certified as a solutions architect on AWS and on Azure, and the control frameworks that decide whether a system is allowed to go live at all.
Omnia is the evaluation pipeline that got rebuilt once too often when the runtime changed, generalisedgeneralized, and the answer to who authorisedauthorized this call was one service account too many.
The definition is public. The runtime is open source.
What an agent is lives in PromptPack, an open specification. What executes one is PromptKit, open source. The four deploy adapters are public repositories. What you write stays in your repository and runs on code you can read, and Omnia’s always-free set runs in production for good.
Point it at an agent you already run.
Or bring the hard question. Either way you get the person who built it.