A capable agent can still be a weak product if the user has to invent its first job. Browserbase’s interview with Gusto co-founder Eddie Kim calls this the blank canvas problem: install OpenClaw or open a powerful coding assistant, then fall back to search because the product has not connected capability to a recurring responsibility.
Gusto’s answer is Cofounder, an agent built around work the company already understands. Small businesses run payroll on a schedule, approve time off, administer benefits, and handle recurring HR tasks. Those workflows already have data, timing, exceptions, and decision owners. Cofounder can propose a concrete automation, run it on schedule, and interrupt the owner only when judgment is needed.
This is a different product contract from a general chat box. The starting point is not “what would you like to do?” It is evidence that the customer performs a known task repeatedly. The product turns that evidence into a suggested delegation with an explicit schedule and human boundary.
The development story reinforces the same point. Browserbase reports that four engineers and one designer built Cofounder in ten weeks, producing roughly 1,000 pull requests with a nine-minute median review time. Kim’s emphasis is that the coding agent became productive through concentrated feedback from the team. Capability was available; the scarce input was teaching the system what good work looked like.
For New Runtime, this is a useful adoption test. Before adding another autonomous surface, identify one repeated job with a stable trigger, observable completion state, and named decision owner. If the system cannot propose that job from existing operational context, the user is still holding the blank canvas.