How Much Autonomy Should AI Agents Get?

Pooja Brown says the first trade-off, made every day, is how much autonomy to give the product on day one. Everybody wants fully autonomous versions running immediately, but a team earns the right to full autonomy one run at a time. She says they can aim to build a fully autonomous agent and still have to bring people along. The agents have to understand the work people already do, create output, create value, get corrections from the team, and then keep compounding on the corrections. The level of autonomy to give agents is a decision made constantly.

She learned the hard way while replatforming things at DocuSign. One of the big things the system did, she was responsible for transforming. It was critical to understand what the old system did and how it worked, then bring people along to the new system even if it was a lot better. So many workarounds and bugs were built into the old system that they needed to port them over as new capabilities, new features, and new ways of working. She also weighs how much autonomy to bring into the systems, and the threshold for when agents displace a meaningful amount of work for humans. Solving a client's problem from scratch is much faster right now. Building reusable pieces is slower, but the pieces compound across more use cases for the customer and create more value. The slower path takes tension and conversation.

She copes by naming the trade-offs out loud and bringing her teams and the customer's teams along, so they know why some things slow down to create value later.