Run paid search with proposing agents, a decision engine and simulated changes

Josh Odenheimer on Slack's growth team rebuilt how the company runs paid search, using a set of agents over a SQLite data store. Discovery, copywriting, strategy and optimisation agents can only propose changes, and a decision engine weighs those proposals against each other. Before anything ships, the plan is tested in simulation over the previous 30, 60, 90 and 180 days, then goes through pull request review, and the team gets a nightly digest in Slack. Comparing 12 weeks with the agents against the previous 12, cost per acquisition on the main conversion fell 18% even as spend went up, and weekly time spent managing campaigns dropped from eight to ten hours to about one.