Services / AI agents
We build agents as engineered systems: a model in a loop with scoped tools, memory, guardrails, evaluation and a budget. You choose the model, and you can swap it without rebuilding anything. A core team leads, and dedicated engineers join where the work needs them.
01 / The loop
An agent is a loop around a model, never a single prompt. Every pass through the loop is bounded, logged and checked.
02 / Building blocks
The same parts appear in every agent we build. Each one is configured, tested and owned, so you can see what the agent is made of.
03 / Your own model
The harness talks to models through one interface. You choose the model per task, and changing it is a configuration change: prompts, tools and evaluation stay where they are.
Your choice per task
| Option | Where it runs | A good fit when |
|---|---|---|
| A provider API | The provider's platform, reached through the gateway | You want the strongest general model and the data may go there |
| A model you already license | Your existing agreement, connected and governed in the gateway | A contract is in place and the task is covered by it |
| An open-weight model | Your own infrastructure: Nebius in European regions, on-premises or Azure in an EU region | Data must stay where you decide, or volume is steady and high |
| A fine-tuned model of your own | Served on endpoints you control | The task is narrow and a smaller model can do it well |
The model is one line per task. Everything around it stays.
# Schematic configuration, not real syntax
tasks:
classify_request:
model: small-model-a
draft_reply:
model: large-model-b
extract_fields:
model: our-fine-tuned-model
gate:
evaluation_set: real-cases
quality: must-hold
cost_per_task: within-budget
rollout: shadow-run, then traffic-split
04 / Patterns
We pick the simplest pattern that does the job and add structure only where the task needs it.
| Pattern | What it does | When to use it |
|---|---|---|
| Single tool-using agent | One model calls a small set of tools in a loop until the task is done | One well-defined task with a few tools, the place to start |
| Router to specialised agents | A first step classifies the request and hands it to the agent built for it | Several distinct request types that each need their own tools and rules |
| Supervisor with sub-agents | A supervising agent splits a task, delegates the parts and combines the results | Tasks that decompose into independent parts with different skills |
| Planner and executor | One step writes the plan, another carries it out step by step | Long tasks where the plan should be reviewed before anything runs |
| Agent with human approval | The agent prepares an action and a person confirms it before it happens | Any action with consequences, such as a booking, a payment or a message to a customer |
| Multi-agent workflow as a graph | Nodes and edges define the flow, with checkpoints that save state at each step | Processes that must resume after a failure and be audited step by step |
We build these with the frameworks your team prefers, for example LangChain, LangGraph or PydanticAI, with MLflow tracing and tool access through MCP.
LangChain
MLflow
OpenTelemetry
DatabricksOmnigent is an open-source meta-harness from Databricks that sits above the agents you already use, such as Claude Code, Codex and Pi, or agents you write yourself. It gives them one interface, so the agent and the model underneath stay swappable. Policies for permissions and cost are enforced outside the prompt, and model access runs through Unity Gateway.
Open source in alpha; managed version in Beta. Last verified: 11 November 2026. Our rule: preview features run in a test environment until they are generally available.
05 / Tools and MCP
Tools are where an agent touches your systems, so they are the most carefully governed part.
06 / Memory and state
Memory is a design decision with a retention rule, not a feature that happens by default.
We store state in Lakebase, the managed Postgres service on Azure Databricks, governed in Unity Catalog next to your lakehouse. Current-state data in small reads and writes is what Postgres does well, and pgvector is among the supported extensions.
OLTP, OLAP and Lakebase in depth
Databricks
Lakebase
PostgreSQL
Unity Catalog07 / Evaluation and safety
An agent is only as good as the evidence that it works. These controls are part of every build.
08 / Running it
After go-live, an agent needs the same operations as any service. The Unity AI Gateway below is where access, spend and logs are governed.
Governance for AI
Every call from teams, agents and coding tools passes through Unity AI Gateway, so access, spend and logs follow the rules of the rest of your data.
Generally available. Smart Routing is in Beta. Last verified: 10 November 2026. Unity AI Gateway documentation
09 / Approach
Five steps, each ending in something you can review. Everything is code in your repositories.
One task, one owner and one number that tells you whether the agent does it well.
The test set, the tracing and the regression gate exist before the agent does much.
Each tool arrives with its service account, its tests and an approval step for writes.
The agent proposes and people decide. We compare its proposals with their decisions and tune.
Only the cases where measured quality is good enough move to lighter review.
10 / Proof
We set up AI harnesses and frameworks for clients, and we build every evaluation set on your own tasks.
At RIVM, the Dutch National Institute for Public Health, we built LLM operations frameworks and multi-agent proofs of concept with MLflow, LangChain and Azure AI Foundry. Read the RIVM case
We completed the Databricks Delivery Expert training track for Agent Evaluation.
AI workers: role-based agents for operations ML and LLM operations AI engineering solutions
11 / Deliverables
Each starts with a conversation about your scope.
Typical duration: 2 to 3 days
In 2 to 3 days we map how data and documents flow today and rank the three AI or data use cases with the best value for effort.
What you get
Typical duration: 4 to 8 weeks
We build the path to production for your models and LLM applications on Databricks: registry, pipelines, evaluation gate, serving and monitoring, with your first use case running through it.
What you get
Further reading
Written by the engineers who build these systems.
AI Engineering 13 min
Build trace-driven model routing with task-level benchmarks, quality gates and cost-per-success measures, without assuming Azure or Databricks integration.
Read articleDatabricks 10 min
Plan a governed Databricks Agent Memory pilot with retention rules, repeat-session evaluation, user-isolation tests and clear deployment promotion gates.
Read articleAI Engineering 11 min
Govern coding agents with Unity Gateway: approved models and tools, clear authentication boundaries, trace access, spend ownership and EU processing checks.
Read articleFree check
A score from 0 to 100 and a next step, in about three minutes.
12 / FAQ
A chatbot answers questions. An agent works toward a goal: it plans, calls tools in your systems, reads the results and checks them before it answers or hands over. That is why an agent needs scoped access, evaluation and a budget.
Only when you decide so. We start with proposals that a person approves. Cases where the measured quality supports it move to lighter review, and every action is logged and can be reversed.
The ones you choose. The harness works with provider APIs, models you already license and open-weight models on your own infrastructure, and a model can be swapped by configuration. Each model is tested on your evaluation set before it handles a task.
In layers. Agents run with least privilege, so an injected instruction can reach only what the role allows. Write actions wait for approval, and guardrails check tool calls and answers. We also red-team with injected documents and tool output before go-live.
Cost depends on tokens and calls. We set a budget per agent through the gateway and measure cost per task, so you see it before and after every change. The page on cost and token optimisation shows how.
Next step
Tell us about one task, the tools it needs and the model you would like to use. We will say what the first engagement involves.
Currently accepting projects for Q4 2026