Agentic AI

AI agents that work inside precise boundaries. Decisions stay with people.

Agentic AI

From assistant to agent

An AI agent does more than answer. It plans, uses tools, consults documents and systems, prepares drafts and checks and repeats the cycle until it reaches the goal. It is the step from an assistant that suggests to a system that works. What matters is not only what the model can do, but also what the agent is allowed to do.

We design agents and multi-agent systems for companies and public bodies with one rule: the agent acts, a person governs, inside a boundary the system cannot cross.

How we design agents

Boundaries before capabilities

Every agent has minimum permissions, budgets and a perimeter it cannot cross. CyberAgent, for example, only scans domains the customer has proved it controls.

The agent proposes, a person decides

Agents read, prepare and propose. Actions with impact (sending, signing, changing data) go through a person’s approval.

The harness matters more than the model

An agent’s behaviour depends largely on the harness: the context the model sees, the loop with its limits and exit conditions, the graph of agents and tools. The model can be swapped, the harness stays.

Local, with open-weight models

When data cannot leave, agents run on site with open-weight LLMs, including on NVIDIA DGX Spark, on the customer’s premises or in a cluster.

Tools connected through MCP and A2A

Agents use tools and data sources through the Model Context Protocol and talk to each other with A2A, with explicit permissions for every tool.

Measurable governance

Policies, personal data redaction and audit trails on every model and tool call, following the OISG paradigm: every pillar has its own metric.

Governing agents

We adopt OISG (Open, Intelligent, Secure, Governed), the paradigm for autonomous AI systems proposed by Stefano Noferi: four pillars, each with a metric, and a 20-criterion adequacy test.

Admina, a project by Stefano Noferi, sponsored by noze, is the open-source framework that puts it into practice: every model and tool call goes through a loop breaker, prompt injection filtering, personal data redaction, policies and a SHA-256 hash-chained audit trail. DataGovern, IntelliPA, AgenticHealth and VerdePilot are built on Admina.

With DebugABot we research runtime controls for autonomous agents: kill switches, fingerprinting, alignment verification and deception detection.

Frequently asked questions

What is Agentic AI?

It is AI that does more than answer: an agent plans, uses tools, reads and writes data and repeats the cycle until it reaches a goal. An agentic system can combine several specialised agents that hand work to each other.

What is the difference between a chatbot and an AI agent?

A chatbot answers a question. An agent works on a task: it consults documents and systems, calls tools, prepares drafts or checks and stops when the goal is reached or when a human decision is needed.

Do AI agents act on their own?

Only inside the perimeter we define. In our systems agents read, prepare and propose; actions with impact (sending, signing, changing data) require a person’s approval.

Can Agentic AI run locally, without the cloud?

Yes. We use open-weight LLMs run on-premise, including on NVIDIA DGX Spark, so documents and data stay inside the customer’s network.

Which frameworks do you build agents with?

With the main agentic frameworks (LangGraph, CrewAI, Microsoft Agent Framework, LlamaIndex, smolagents, Google ADK) and with custom harnesses, connecting tools and data through MCP and A2A. We choose the framework based on the problem.

How do you govern a system of agents?

With explicit boundaries, human approval of actions with impact and checks on every call: policies, prompt injection filtering, personal data redaction and a verifiable audit trail. We follow the OISG paradigm and use the open-source Admina framework.

Contact us

Have a process to hand over to agents?

Fill in the form: your request goes to the dedicated team.

Contact us now →

Need support?Under attack?Service Status
Need support?Under attack?Service Status