Working Surfacelink · 18 Nov 2026
03
A mail-sorting machine dropping hundreds of letters into rows of pigeonholes while a few odd envelopes slide onto a small side tray, where a figure seen from behind reads each one before placing it, the whole scene drawn with wide empty margins on every side.
Salesforce Engineering

How AI Slashed 600-Test-Failure Triage From Days to 90 Minutes

Shireen Nagdive and Riya Suri · 17 November 2026

Company story · agents · decisions · Human approval

Read the original
Takeaway

A Salesforce engineering team cut triage of over 600 failing tests from days to 90 minutes, with code doing routine steps and engineers approving the AI's conclusions.

Summary

Shireen Nagdive, an engineering manager at Salesforce, leads a team retiring an old login method for the SOAP API, an interface other software uses to connect to Salesforce. Enforcing the change broke more than 600 tests owned by several teams, and real faults looked the same as the expected breaks. Nagdive and Riya Suri describe the tool the team built with Claude Code to sort and route those failures.

Key points
  • The tool gives Claude Code or Cursor eight defined actions on GUS, the internal Salesforce system that tracks work items and test failures.
  • Ordinary code fetches failures, groups them by error pattern and reassigns them, while the AI model reads unfamiliar errors and unclear ownership.
  • Engineers review the tool's conclusions and proposed actions, and keep the authority over where work is assigned.
  • Triage of the 600 failures was estimated at several engineers for two to three days, and took one engineer about 60 to 90 minutes. Three or four teams have used the tool.
Implication

An engineering team facing many similar failures can leave fetching and routing to ordinary code. It can use an AI model only for unclear cases and keep a person approving every move. The team has no usage data yet beyond feedback in Slack.

Plays
Engineering

Let ordinary code fetch, group and reassign test failures, use an AI model only for unfamiliar errors, and have an engineer approve each conclusion.

Engineering

Write a specification of the problem and the required behaviour before asking a coding agent to build an internal tool.

Derived by Working Surface from the article; more in the Playbook. Source line: Deterministic code handles API calls, identifiers, validation, and reassignment mechanics.

Source issue

18 November 2026: when AI agents make a team's output steady, the team's attention moves to the places where work stalls and to the people who check it.