04 / 08 · Service
Processes that read the world through tools and write their action back.
§ Manifesto What we actually do
The part nobody reads. The part that matters.
An AI agent is not a chatbot. It is a process. Patient, stateful, scoped. That reads the world through tools and writes its action back. We build agents the way we build good cathedrals: with small, sturdy decisions and one clear purpose.
We choose Claude where judgment matters, GPT where coverage matters, and open-weights where the job must stay inside your walls. We write the tool-calls ourselves. We log everything. Prompts, completions, retries, refusals. We decide how it fails, because failure is the only part of the contract you cannot outsource to a model.
Good agents have memory that degrades gracefully, budgets that they cannot overrun, and a voice that belongs to your brand and not to a foundation lab. We keep humans in the loop where humans matter and keep them out of the loop where they become bottlenecks.
The test of an agent is not a demo. The test is whether, three months in, it is still doing the job, and whether the person whose job it used to be has moved on to something only a human could do.
§ Stack How we build it
Tools chosen, not inherited.
§ Evidence Things we have actually shipped
§ Next If any of this resonates
We answer real questions in a real window of time. Hello always beats a form.