Practical writing on building and operating AI agents at scale — architecture, channels, retrieval and how agencies run agent fleets for their clients.
One agent is a project. Three hundred across twenty organisations that must not see each other is an architecture, and the difference is not scale. Five boundaries that decide whether a platform can be operated as a fleet.
Personal data in LLM flows: redacting it from the prompt protects the model and breaks the tools. How AI Agent Builder swaps it for stable tokens on the way to the model, gives the real value only to the tool that needs it, and shows the customer their own words on the way back.
A scheduled agent that labels every incoming email with a System One decision model, archives the noise and leaves only what needs a person. How AI Agent Builder puts the model in exactly one step, why everything around it is ordinary workflow, what a run tells you, and what it costs per email.
Most of what an agent does is decide — which category, whether it is in scope, which tool. How AI Agent Builder runs those decisions on Jev, TypeSafe's System One model, which answers with a value and a confidence instead of text: its own slot, a fallback you chose, a stated path for every uncertain answer, and a record of who answered.
What a ten-minute onboarding looks like from the operator's seat: what the client inherits instead of receiving, what you cap before they start, and the test that decides whether their agent goes live at all.
Three questions about your own fleet that most operators cannot answer, the record that makes them answerable, and why retirement is the lifecycle step that nobody builds.
The gap between ten minutes and three days is waiting on people, not labour. How AI Agent Builder puts a new client on the platform with no person in the middle: the tenant boundary, limits written as values, connections resolved and never copied, an agent imported from an artifact, and where each guarantee stops.
The user's question is under one percent of what you pay for. An arithmetic model of fleet inference cost denominated in tokens, the four multipliers that hide inside it, and the one term that grows quadratically.
Provider outages last hours, not minutes, and a fallback chain three deep can still be one provider. A teardown of what actually has to be in place, minute by minute, and what failover costs you when it works.
Prompt injection is one entry point out of five, and the input filter everyone installs guards exactly one of them. Seven manipulation classes, why the obvious defence fails in each, and where the check has to sit.