The newspaper of the Forward Deployed Engineer
Tools FDEs use every day to deploy, integrate data and evaluate AI at customer sites.
Installing Metabase takes minutes. The hard part is the three decisions that follow: the database account, group permissions and the embedding method. Get any one of them wrong and a polished afternoon demo turns into next week's incident.
09/10/2026
In Pydantic AI, switching models takes a one-string edit. To show a client the agent still works afterwards, you also need an eval suite and a pinned library version.
At a customer site, you can only trust an answer to "is the new prompt better?" if a score runs on every pull request. Braintrust is built for exactly that job.
With drag-and-drop, you can have a customer demo running very quickly. Whether that demo can run in production depends on three things: the limits of the Code node, the scope of the licence and how the system scales.
If you have to install one product into five customers' Kubernetes clusters, each with its own configuration, you need a package that is versioned, keeps a history and can be rolled back with a single command.
On a client site, an FDE rarely gets to choose the ETL tool. Usually the job is to take over a pipeline that is already running, so the first task is to understand it before changing anything.
At a customer site, a good model is not enough. You also have to show which version is running in production, where it came from and how to rebuild it.
At a client site, the hard question is rarely whether the agent can do the job. It is who gets to press the button that lets it, and LangGraph's interrupt function was built for exactly that moment.
Clients will ask why the system answered as it did, what it costs and whether it is getting better. One open-source tool puts all three answers on a single platform.
