Salesforce’s agents have completed 7 billion Agentic Work Units (AWUs), with 3.2 billion in Q2 alone. The company overcame early reliability issues by moving from a single-prompt design to a multi-layer system combining deterministic code with reasoning models. Leaders Muralidhar Krishnaprasad and Khushwant Singh explain how Salesforce built a robust agent stack including identity resolution, context engines, and observability tools.
The article details why naive prompting fails beyond eight instructions—models treat rules as context, not mandates. Salesforce now enforces behavior through code, data layers, and evaluation pipelines, ensuring agents handle sensitive tasks like job interviews and patient data reliably.
Source: ByteByteGo · Summarized by HeadlinesBriefing