From a paragraph describing boring work to an agent that does it
Someone describes the repetitive work on their plate in a few sentences. That description has to become a working agent, with boundaries, without a month of discovery.
The bottleneck in agent work is never the model. It is turning a vague description of a job into a process definition precise enough to automate: the trigger, the inputs, the decisions, the exceptions, and what done looks like.
We automated the first draft of that. The intake reads the description, or an uploaded job posting or SOP, and produces a candidate process map with the decisions and exceptions made explicit, including the ones the writer did not mention. A consultant then sits with the customer and corrects it, which is a far shorter conversation than starting from a blank page.
From the corrected map, the agent is scaffolded: the integrations, the guardrails, the human sign-off steps and the logging. The consultant watches the first runs against real work before anything is trusted.
What runs now: intake to draft process map, consultant correction, scaffold to a running agent with sign-off steps and logs, then monitoring with the consultant named on the account.
What it taught us: automating the discovery step is worth more than automating the build. A correct process definition is most of the work, and it is the part clients cannot easily do alone.