Agents that remember, search and prepare, people who decide. Here is what works today on the sites published by QDNA, what still needs watching, and how we help you frame it in your own organisation.
Each block removes a repetitive task from the person doing the work, without taking the decision away. For each one, in a few lines: what it does, what already works on our own sites and the limit to keep in mind.
The memory combines three layers: a knowledge wiki, a hybrid recall that pairs vector search with a relationship graph, and a per-project memory loaded when each session starts, so the agent picks up where it left off.
What works. The agent finds durable facts without being told again, and the embedder and reranker both run locally. Limit. A memory is consolidated and reviewed: it complements your reference sources without replacing them.
An orchestration daemon written in Rust, save-agent-rs, coordinates specialised agents behind the LiteLLM gateway. For a factual question it first looks in a dated fact graph, then queries several models and keeps only the answer they agree on.
What works. Every verified fact is kept with its date and provenance, so the same question is not asked twice. Limit. Consensus reduces errors without removing them, and each check costs model calls.
The augmented organisation's digital twin mirrors the company's org chart. Each business role has its own agents, budgets and partitioned access rights, and the control tower shows live who is working, who is waiting and what needs a decision.
What works. Any sensitive action waits for a human sign-off before it goes out, and idle roles go to sleep. Limit. The demonstrator runs on fictitious data: each organisation sets its own roles and thresholds.
Monitoring and document research agents query the web, read pages and find documents and videos by their visual content, through MCP connectors hosted on site, with no dependency on a third-party service.
What works. Every summary comes with its sources, which you can open and check. Limit. A source has to be read: the agent prepares the summary and flags its references, the person validates what it cites.
The agents rest on a simple stack, read from top to bottom. Each layer relies on the next one and has its own sheet, from the agents down to the hardware that runs them, on premises or hosted.
Each layer is chosen, and the hardware sized, for the target concurrency, throughput and context. The first estimate comes from the memory calculator, then gets checked against your real workload.
Three habits are enough to get started.
To learn step by step, the free course Learn AI (external site) starts from the basics and goes up to advanced uses.
Hermes Agent is the agents' execution engine.
OpenCode orchestrates software development.
The harness is the execution loop that turns a model into an agent.
LiteLLM exposes a single interface in front of every model, local or remote, making them interchangeable.
The semantic router classifies every request before LiteLLM.
A set of self-hosted MCP connectors equips the agents: memory, visual document search, wiki recall, web search and reading, and more.
The memory combines three layers.
For the logic that decides, request by request, which model handles what and what leaves the company, see the guide AI orchestration engine: routing every request to the right model.
It is the set of building blocks that turn models into agents. A harness runs the agent loop and a LiteLLM gateway makes models interchangeable. A semantic router picks the model request by request, MCP servers give agents their tools, and a three-layer memory completes the whole.
LiteLLM exposes a single interface in front of every model, local or remote, and applications call that one interface only. The semantic router upstream decides, by sensitivity and task, which model answers.
The semantic router classifies each request before any model call. Confidential content stays on a local sovereign model and only mundane content may overflow to an external API. The organisation's policy is applied before execution rather than after the fact.