The rise of the Agent Builder

For companies to pull ahead this next year, they need to systematically replace SaaS tools and manual workflows with agents, and dedicate a full-time role to do exactly that. We call this the Agent Builder. Our first Agent Builders are Emily, who started off building agents to automate work on the Growth team, and Dave, who was building agents for product intelligence. Their mandate is to use our cloud agent platform Oz to automate tasks. We weren’t sure about this at first since Oz is built around coding agents… but it turns out that a great coding agent makes a great general purpose agent. These agents complete tasks that we might have purchased a SaaS for a few months ago – e.g. social listening and user feedback analysis. Indeed, we are cancelling SaaS subs we had for these tasks in favor of rolling our own agents. We've already automated: Fraud detection: an agent that constantly looks for suspicious usage patterns and creates PRs to block users (and also identifies novel fraud patterns). This is saving us tens of thousands of dollars every day. Competitive intelligence: agents that monitor competitor launches, compare them to ours, and post weekly market summaries to Slack — replacing what used to take our product team half a day each week. Enterprise POC summaries: summaries of how our individual enterprise pilots are performing, notifying us where we might want to have a Sales Engineer step in and help so that the pilots are successful. Everywhere we look, we see new use cases for agents. That’s why we decided to make building them a full-time job. Our old internal mandate was "start with a prompt" to get help with a task. Now, the mandate is to see if the task can be automated on an ongoing basis by building an agent. Building these internal agents presents a bunch of challenges. There’s up-front environment setup, data access grants, monitoring, and prompt tuning. That last point is best served by dogfooding with our team early, forgiving the kinks, and turning feedback into prompt improvements. We are also discovering new primitives for our Oz cloud agent platform to enable these workflows (here’s an early demo of orchestration). Overall, it's never been easier to automate a tedious task without a structured, Zapier-like workflow. An agent, given the right capabilities and skills, can use intelligence to decide on the right workflow instantly. You don't need to predefine a series of "if this, then that" steps. What you really need is high-quality auditing, steering, and tracking mechanisms to see how these agents perform. You need the tooling to improve them over time. And in the future, these same primitives will allow agents to start improving themselves. It sounds like science fiction, but it’s not – agents can look at their past runs, see what could be improved, and change their own definitions. It’s going to be a wild ride over the next year. To figure out the future, we’ll need Agent Builders at every level to make the most of this technology. My prediction: within a year, every company over 50 people will have at least one person whose full-time job is building internal agents.