The integration question
Getting your data in is no longer the expensive part
Every data platform evaluation runs into the same question: how do we get our data in? For years the honest answer was a quarter of roadmap and a senior engineer per system, and it is the answer that stopped most projects before they started. That estimate is out of date. Integration is mapping work, and coding agents are very good at mapping.
Why it got cheap
Integration is plumbing, and plumbing is what agents are best at
You understand the contract at the source, the contract at the destination, and you write the mapping between them. It has always been more code plumbing than computer science, which is exactly why an agent does it well.
The instrument side is standards
OPC-UA, Modbus and the rest of the plant protocols are standard work with a standard answer. Getting readings off a control system is not a bespoke project, and it never needed to be.
The business side is mapping
ERP, maintenance, permits, spreadsheets. Read one contract, read the other, write the transform, test it against real rows. Mechanical, checkable work, and an agent does not get bored on field four hundred.
Failures come back readable
A wrong mapping throws, or produces a value that fails a check. An agent reads that, corrects it and runs again in the same sitting. That loop is why the second version costs almost nothing.
Where the constraint went
An agent can build anything. It cannot guess what your pump is called.
So the data lands. That is the easy half. Point an agent at your historian and it will happily write code against it. Ask which of four thousand tags feeds the emissions figure in last month's report, and it has nothing to work from. PT-4021 is a string. That it is the discharge pressure on the pump behind that number lives in one engineer's head.
So every agent you build asks the same question again, and every one of them gets a slightly different answer. That is the cost of doing nothing, arriving with every new agent.
Moving the data is nearly free. Meaning is not.
What makes it actually free
Give the agents something solid to build against
One model of your operation: the assets you run, the functions they serve, the business context that makes a number matter, and the signals and events they produce, all linked. Agents reach it through the operational data platform, and every call carries the permissions of whoever made it.
Build once, not once per agent
Every agent you add inherits the same definitions, so the second application is cheaper than the first and the tenth is nearly free.
Bring your own model
Anthropic, anything OpenAI-compatible, or an endpoint on your own hardware. We build the agents with you and set the policies they run under. The language model behind them is your choice.
Let them reason, keep them fenced
An agent does not replay a script. It reads a workflow the way a new engineer reads a runbook, works through the steps with the model as its context, and adapts when reality disagrees with the plan. Policies draw the fences: what it may read, what it may change, and where it stops to ask a person.
How we help
Engineers next to yours, until the model is real
We are deliberate about what we sell. We do not sell a connector marketplace, and no platform, ours included, makes the modelling decision for you.
Two weeks on site
We watch the actual work, rank what an agent could take over by hours and cost, then build and deploy one working agent against your data. You keep the code and a measured baseline.
Embedded engineering
Ontology mapping, iterative development, deployment and training. The people you meet in week one are the people still there in week four.
Expertise that stays
Core is open source under AGPL-3.0; the SDK clients are Apache-2.0, so your own code is never touched by copyleft. The point of an engagement is that you can do the next one without us.
And afterwards
The agents keep it running too
An API changes, a call fails, and the failure comes back as something an agent can read and repair. Maintenance that used to consume a team is increasingly work the same agents absorb.
Which is the second reason to get the model right. It is what tells them whether the repair was correct.
Common questions
Asked and answered
How do you get data into a data platform?
Through integrations. On the instrument side, standard protocols like OPC-UA and Modbus move readings off control systems. On the business side, mapped transfers bring in ERP, maintenance and permit data. Each integration reads the source's contract, maps it to the platform's model, and moves the rows.
What does a data integration cost today?
Far less than the estimate most roadmaps still carry. The plumbing, connectors, field mappings and transformations, is well-specified work that coding agents produce in hours rather than weeks. The cost that remains is deciding what the data means, which pump, which process, which question, and that work belongs to the model, not the pipe.
Do AI agents really write integration code?
Yes. Integration is contract-to-contract mapping, some of the most well-specified work in software, and randomized studies show the largest agent gains on exactly that shape of task. The agent reads both contracts, writes the transform, tests it against real rows, and corrects its own failures in the same sitting.
What happens when a source system changes its API?
The call fails, and the failure comes back as something an agent can read and repair. Integration maintenance that used to consume a team is increasingly absorbed by the same agents that built the integration, which is also why the model matters: it is what tells them whether the repair was correct.
From data mess to data success
Ten minutes to a working instance with sample data. No sales call required to find out whether it fits.