In 2018 we put an AI agent on a construction worker's phone. Not a chatbot that answers questions, an agent that does a job: it asks a foreman the right questions about the day's work, reasons about the hazards, writes the Pre-Task Plan, and files it back to the project's system of record. All over a text message, with no app and no login.
At the time we did not call it "agentic." Nobody did. The word the industry now uses did not exist in common use yet. We just called it Nyfty, and we shipped it to real crews on real sites.
What "agentic" actually means in production
The current excitement about AI agents tends to skip the boring parts. The boring parts are where the value is. An agent that does real work has to:
- hold a goal across a multi-step conversation, not just answer one prompt;
- decide which questions are worth asking, and skip the ones that are not;
- produce a structured output a downstream system will accept; and
- do all of it reliably enough that a foreman in muddy gloves never has to think about the AI.
We learned those lessons in the hardest possible environment: safety-critical work, non-technical users, and zero tolerance for a clunky experience. That is a long way from a demo.
Why this matters for everything we build now
Yakka Labs is the studio those lessons feed into. When we built Backlit, our hosting platform for apps built with AI, the instinct to handle the unglamorous, risky parts for the user, sign-in, data, deployment, came straight from years of shipping agents that had to just work.
We did not arrive at agentic AI because it trended. We were already there, in production, before it had a name. That is the foundation under every product we make.