About
A product born from a real application.
InApp-Agent did not start from a blank page. It comes from BANANA-CONTENT, where we built and then operated an agent genuinely embedded in a complete business application.
What BANANA taught us
Building an agent that answers is relatively simple. Building one that acts inside an existing application, without breaking its rules or misreporting what it did, is far harder.
The real difficulties lie outside the conversation: knowing exactly what the agent is allowed to do, telling a success from a half-success, keeping context when infrastructure restarts, and making sure a deleted preference does not return through a back door.
We met these problems in production. That is what lets us turn this into a product today rather than a demo.
Our design principles
Let the agent reason
An over-constrained agent becomes useless. It must be able to choose what it reads, rephrase and combine available capabilities — within an explicit boundary.
Make permissions explicit
What the agent can do is an architectural decision, never a consequence of what someone types at it.
Report what happened
Prepared, executed, partial, refused, uncertain: these states are distinct and stay distinct all the way into the interface.
Improve with evidence
A change is justified by a finding and verified by an effect. Vigie links the two.
Legal entity information, contact details and the team will be published on this page before the site opens publicly.