Session abstract
What you’ll learn
Most AI agent projects fail to launch for the same reason: trying to go from zero to full autonomy in one leap, without building credibility to justify the risk. PGA took the opposite approach while building Divot, our internal product development agent. Divot started as a narrow automation for a chore nobody wanted to do, and every expansion since then — from cleaning up feature flags, to reviewing pull requests, to operating as an interactive teammate across the team's full set of tools and context — followed the same cycle: find a bounded problem, pilot it on a small scale, and iterate towards more autonomy as the agent proves itself capable and your company's risk appetite allows. This talk walks through Divot's evolution chronologically, showing the specific problem, intervention, and trust-building evidence at each stage, plus the deliberate ways the team constrained Divot's autonomy early on so they could course-correct before scaling up. By following this strategy, you'll learn to apply AI in impactful ways that automate and elevate the work of your team.