Operators who became platform builders.

Agentsia started as an operator building AI agents, not a platform company. Our first product was a health and nutrition advice agent: frontier model, RAG pipeline, carefully engineered prompts. It worked in demos. It did not work in production.

The failure mode was consistent: the model was good at sounding authoritative but unreliable on the specific dietary and clinical questions that actually mattered. RAG improved recall. Better prompts improved tone. Neither fixed the underlying gap between what a generalised model knows and what a domain expert expects.

We moved the agent into programmatic advertising, where the domain is narrower and the ground truth is less ambiguous. The same problem: frontier APIs were too slow for real-time bidding, too expensive at auction scale, and sent our bid-request payloads to a third-party network.

The solution we needed did not exist. We built it. The tool we used to specialise our own models became Modelsmith. The operational pattern we developed for running training campaigns became the iterate loop. The governance process our team used for approving promotions became the promotion state machine.

We are now making that platform available to other operators facing the same constraints we faced.