Too much access
Agents are often given blanket rights to file shares, mailboxes and business applications - more than the task requires. Every surplus permission is an attack surface that a manipulated agent can exploit.
AEGIS by oQueo - governance for AI agents
Autonomous AI agents. Without losing control.
AEGIS separates thinking, checking and acting. No agent takes action until an independent authority has approved it - run on your own infrastructure, fully traceable, under your control.
Launching in 2027. A limited number of pilot places for selected organisations.
Three authoritiesPlanning, checking and execution run as technically separate components. None can override another.
Default-denyOnly what a rule explicitly permits is allowed. Everything else is denied.
SignedA step is executed only if it carries a valid cryptographic approval from the checking authority.
LocalRuns on your hardware, even without an internet connection. Your data stays in-house.
The problem
AI agents already act on their own
They read files, write emails and call other systems. Most solutions on the market simply trust the model to do the right thing. Yet a language model cannot reliably tell your instruction apart from a piece of text it happens to be processing.
Acting without approval
What an agent plans, it often carries out in the same step. A single injected instruction in an email or a PDF is enough to turn a summary into an outbound transmission. No second, independent check takes place.
No traceability
Who triggered what, when, and on what basis? With many systems this is almost impossible to reconstruct after the fact. Without a reliable log, you lack the basis for root-cause analysis, liability questions and evidence for supervisory authorities and clients.
The principle
No agent checks itself
AEGIS is built on the three-authority principle. The work is split across three technically separate authorities. Each has exactly one job - and none can override another.
-
Think
Planning authority
Understands the task and plans the steps. The result is a structured proposal known as an intent. This authority has no access to credentials and cannot execute anything itself.
-
Check
Checking authority
Matches every intent deterministically against your rules: permitted tool, permitted action level, valid timestamp. Only what passes every check receives a cryptographic signature.
-
Act
Execution authority
Verifies the signature a second time and executes only what has been approved. Every step, including every denial, is recorded in a log that cannot be altered afterwards without detection.
audit.logExample
- 14:22:31planintent 7f3a read_document contract.pdf level 0proposed
- 14:22:31check7f3a tool permitted level 0 signedapproved
- 14:22:32act7f3a signature valid executedlogged
- 14:22:35planintent 9c1e send_mail external level 2proposed
- 14:22:35check9c1e level 2 above permitted maximumdenied
Use cases
Who AEGIS is for
AEGIS is designed for organisations that want to use AI agents in production, whose data must not end up in third-party hands and whose decisions have to be demonstrable.
Law firms and tax advisers
Client data is subject to professional confidentiality. Agents may research and prepare drafts, but nothing leaves the firm without being checked.
Healthcare
Health data is one of the special categories of personal data under the GDPR. Local operation and strict action levels keep sensitive data in-house.
SMEs with evidence obligations
If you owe evidence to clients, auditors or supervisory authorities, you need a log showing which agent did what, with which approval.
Teams that want to stay independent
AEGIS is not tied to any model provider. You choose the models that suit your requirements and replace them without rebuilding your controls.
Data sovereignty
Your data stays with you
AEGIS runs on your own infrastructure, from an on-premises server to an environment with NAS storage. No disclosure to third parties, no dependency on any particular provider.
By default, the internet is treated as untrusted. You decide which models are used, which data the system is allowed to see at all, and whether it may reach destinations outside your network.
- Works offlineThe core operates without an internet connection, and models run locally.
- No training on your dataDocuments, logs and the knowledge base never leave your infrastructure.
- Free choice of modelsIndependent of models and providers, with no vendor lock-in.
- Storage location of your choiceLogs and the knowledge base reside on your server or NAS.
Insights
Insights on AI agents
An in-depth article every week on governance, security and law relating to AI agents - with sources and without marketing language.
AI agents and the AI Act after the Digital Omnibus: roles, duties, deadlines
Which AI Act duties apply to AI agents, and from when: provider or deployer, high-risk classification, AI literacy and the new dates after the Omnibus.
Least privilege for AI agents: permissions, accounts and credentials
How much access should an AI agent get? Its own account, rights scoped to the task, and credentials kept out of the prompt, short-lived and revocable.
Human in the loop: designing approvals that actually protect
Human approval of AI agent actions only protects when it is rare, specific and well designed. What AI Act Article 14 and automation bias research tell us.
Questions
Frequently asked questions
What is AEGIS?
AEGIS (Agent Engine - Governed Intent System) is a system by oQueo that lets organisations deploy AI agents without giving up control. Before any action by an agent is executed, an independent authority checks it against your rules, signs it and logs it.
When will AEGIS be available?
Commercial launch is planned for 2027. Until then we are working with a small number of pilot organisations who help shape how AEGIS is deployed in their environment. Go to the pilot programme.
How much does AEGIS cost?
The scope depends on your infrastructure, the number of agents and your requirements for integration and support. We prepare an individual quotation for each organisation.
Do I need my own servers?
AEGIS is built to run on your own hardware, from a single server to an environment with NAS storage. We work out together in the initial consultation which setup makes sense for your models and data volumes.
Which AI models does AEGIS work with?
AEGIS is not tied to a single provider and is designed for locally hosted models. We decide together which models make sense in your environment.
Does AEGIS replace data protection and compliance work?
No. AEGIS provides technical controls and evidence: permissions based on the principle of least privilege, checked approvals and a tamper-evident log. The legal assessment of your deployment and the organisational responsibility remain with you. More on security and compliance.