The independent daily for Salesforce admins, developers and architects

Curated by Abhinav (opens in a new tab)

Recommended

Deterministic gates let an AI agent approve production changes in seconds (opens in a new tab)

You will see how Salesforce's ROAM agent automates Change Advisory Board reviews by letting deterministic code decide while the LLM only explains.

Why read this

You will also learn the integration, contract-testing and fail-closed patterns that made it trustworthy.

  1. Verify approver identity and role
  2. Check linked passing QA test run
  3. Confirm pull request and approval trail
  4. Route to human if any gate is unverified
  5. Write every outcome to audit table
ROAM's gate verifies each requirement independently, and routes unverifiable requests to a human

Salesforce Engineering describes ROAM, the Release Operations Agentic Manager, an Agentforce-pattern agent that reviews production patches for 18 Salesforce products. Before automation, requesters and reviewers chased approvals across Slack, change records, QA evidence and release calendars, with median turnaround around 24 hours. The team mapped the Change Advisory Board process first, then built ROAM on an internal platform providing event ingestion, durable workflow state and shared clients.

The main technical lesson concerns integration correctness. An early bug read the wrong field in nested Slack replies and silently dropped approvals, so the model received an incomplete picture. To guard against schema drift, the team added contract tests using recorded live payloads that run on every pull request, and a mismatch fails the build. The core design rule is that the gate decides and the LLM narrates: deterministic code verifies the approver’s identity and role, the linked test run, the pull request and the approval trail, and the model cannot produce or overturn the verdict. Output validation re-reads the change record and fails closed when an explanation contradicts the evidence.

When a gate cannot be verified, for example because of missing links, expired tokens or outdated rosters, ROAM routes the request to a human, which happens for about 8% of requests. Idempotency keys and decision hashes prevent duplicate posts and changing verdicts on retries, and every outcome is written to a durable audit table.

On AI-assisted development, the article notes that roughly 70% of ROAM’s codebase was drafted or refined with Agentforce-assisted coding, while engineers wrote the remaining 30%, and a separate verifier agent re-executes test suites rather than trusting a fix agent’s self-report. Over about 550 reviews in 10 weeks, the team reports zero incorrect approvals, a median response of about 12 seconds and a 95th percentile under 45 seconds, after which ROAM became mandatory for all production patches.

Read the original at Salesforce Engineering (opens in a new tab)

Worth reading?