We build AI agents that do real work in your company.
And we get your team ready for the agentic world. Behind that are the fields you see in the animation, from quantum physics and game theory to cloud. We can support you in each of them.
What “real work” means
A language model alone talks about the task. Only the software around it lets it get the task done: tools, permissions, memory, and a check by someone other than the author.
The model alone
“The cause is most likely a timeout in the login step. I have increased the timeout, the tests should pass now.”
Sounds finished. Nothing was run.
The same model, properly equipped
- AccessReads the test results and the code it was given access to. Nothing else.
- ToolsRuns the tests instead of speculating about them.
- MemoryKnows what went wrong last week. You explain things once.
- CheckA second agent that did not write the fix reviews the evidence.
- ResultThe change, the test run and the review. Or a plain “could not find out”.
The left side is invented, though you have met the original. The right side is how we build and run our own systems.
Where we know how to build that
From the fields that run in the animation above. Two of them explain almost everything.
An agent that says “done” is holding cards you cannot see.
For years we computed how to play against hidden information: at the poker table, with abstractions, equilibria and collected data. The lesson is simple. You do not trust the face, you change the game. In our systems only what lies open on the table counts: the source, the test run, the evidence.
Left to itself, disorder wins.
That is physics, and it also holds for software written by a language model. Every answer is a draw from a distribution, and without a frame the result drifts. Order costs work: memory that stays, permissions that limit, checks that can fail. Hence the name: Alpentropy.
What we make of it for you
Four things. The first is our focus, the other three are what it stands on.
- AI agentsAgents connected to your systems, with permissions, memory and independent checks, so that they do real work and you can rely on the result.Details →
- Getting your team readyWorkshops and joint work on agentic development: running coding agents in parallel, reviewing them independently and staying in control.Details →
- Software engineeringBackends, cloud and the pipelines that ship them. Python first.Details →
- Data science and machine learningFrom the first notebook to a model that runs in production.Details →
To be honest
So that you know where you stand.
- Alpentropy is new and small. The experience behind it goes back to 2006.
- Agents cannot do everything. If your problem does not need one, we say so.
- What we do not know, we call unknown. We do not guess with a straight face.
- We show what was checked and what was not, in every piece of work we hand over.
Tell us what an agent should do for you. You get a straight answer, even when the answer is no.
[email protected]