Working Surfacelink · 22 Nov 2026
02
A railway level crossing where a heavy barrier arm, worked by gears and counterweights, stays down until every coloured token has dropped into its own slot, while a small loudspeaker on a pole beside it only announces the result to a silhouetted figure waiting with a handcart, the whole scene drawn with wide empty margins on every side.
Salesforce Engineering

How an AI Agent Cut Production Change Reviews From 24 Hours to 12 Seconds

Kamalnath Devarakonda · 20 November 2026

Company story · critique & review · agents · Human approval

Read the original
Takeaway

Salesforce's release team built an AI agent that reviews production changes in seconds, while ordinary code, not the AI model, decides whether each required approval exists.

Summary

Kamal Devarakonda, a senior director of software engineering at Salesforce, described in a company interview how his team automated its weekly review of production changes across 18 Salesforce products. The team built ROAM, an AI agent that collects approvals from Slack and internal records, and let fixed code decide whether each approval was valid. Median review time fell from about 24 hours to about 12 seconds.

Key points
  • Each change needs a passing test run on the target branch, two approving developers other than the author, the named architect and a current cloud leader.
  • Code checks each approval separately, the AI model writes the explanation, and when the two disagree the code's result stands.
  • When an approval cannot be confirmed from the records, the agent hands the review to a person, which happened for about 8% of requests.
  • Across about 550 reviews in 10 weeks there were no incorrect approvals, and after three weeks in production release engineering made the agent mandatory for every production patch.
Implication

Teams giving an AI agent a part in approvals can keep the decision in code that checks records and limit the AI model to explaining it. This is Salesforce's own account of an internal tool built on its own agent platform, and a person still made the final approval click at the time of the interview.

Plays
Engineering

Have code, not the AI model, verify each required approval for a production change, and let the AI model only explain the result.

Engineering

Have an AI agent that reviews production code changes pass a review to a person whenever it cannot find a required approval in the approval records.

Derived by Working Surface from the article; more in the Playbook. Source line: When the two disagree, the model loses.

Source issue

22 November 2026: as AI agents make building cheap, the check before production limits the pace, so teams are redesigning who sponsors early builds, what evidence a change must carry, how agent work is batched for testing and how much one person reviews.