Coinbase details Autopilot quality loop for AI customer support

Coinbase engineers described Autopilot, a pipeline where agents write, test and score versioned support procedures while deterministic services provide repeatable grading. Procedures are expressed as code in a constrained DSL, and shared test contracts let the company compare its bot with partner bots. No quantified post-deployment results were disclosed.

• Support procedures are versioned decision trees written in a DSL
• Agents explore cases while deterministic services grade outcomes
• One test plan runs against both Coinbase and partner bots
• No measured gains in testing time or CSAT were shared