How Salesforce uses Claude Code

Salesforce scaled Claude Code to thousands of engineers and doubled effective output, with fewer bugs per line of code than before.



Engineering leaders from Salesforce explain how they measure AI coding ROI with an "effective output" score built with Stanford researchers, and how an agent maturity curve moved engineers from simple prompts to autonomous agents across the software development life cycle. They also cover a decades-old codebase made headless in weeks instead of a year, a marketplace for sharing agent skills, incident agents that find the right team automatically, and a bug that Claude Opus tracked down in 24 hours after human attempts failed.



More here: https://claude.com/office-hours#salesforce



Chapters

0:00 Making Salesforce headless in weeks, enabling Claudeforce

1:31 Measuring ROI with "effective output"

3:18 The agent maturity curve, from prompts to autonomy

5:53 Driving adoption with live demos and real data

7:30 Building a skills marketplace with evals

9:07 Guardrails: double the code volume, fewer bugs

10:18 Replaying benchmarks to test new models

11:37 "Autonomy without validation is confidence theater"

12:44 How big a task to give an agent

13:52 Biggest mistake: matching model to task

15:13 Advice for engineers and CTOs