Scaling AI Engineering Automation Without Sacrificing Reliability
Salesforce Engineering with Fereydoon Vafaei · 9 November 2026
Read the originalSalesforce's Agentforce engineers let AI agents fix and test their own reasoning engine, but a person must approve every merge and every release to production.
Fereydoon Vafaei is a principal software engineer at Salesforce, whose team uses AI agents to help build the reasoning engine behind Agentforce, Salesforce's platform for AI agents. The team feared that an agent could diagnose a fault, fix it and approve its own fix while being wrong. It answered with a shared record of past incidents, AI models that check one another's work, and a firm rule on what still needs a person.
- The team built a knowledge base that links incidents to earlier fixes, code, Slack discussions and incident notes, and removes outdated material as the code changes.
- Different AI models review each fix so that no model grades its own work, and engineers review the results across many incidents and sign off on them.
- AI agents may design, build, review and deploy changes to test environments, but never merge to the main branch or release to production without a person's approval.
- The team reports that a typical bug now takes one or two days to resolve instead of two to four weeks. One agent can run about ten code reviews at once.
An engineering team can let AI agents act freely where a result can be checked and undone, and keep a person's approval where it cannot. This is Salesforce's account of its own team, and its figures are its own.
Let AI agents build and deploy changes to test environments, but require a person's approval before any change merges to the main branch or reaches production.
When an AI agent writes a bug fix, have a different AI model review it, and have engineers check and sign off on those reviews across many incidents.
Derived by Working Surface from the article; more in the Playbook. Source line: Different models judge each other; engineers sign off.
10 November 2026: as AI agents do more of the building, teams are naming the one point where a person's judgement must sit: approving a merge or release, vouching for a summary or defining good output.