At a glance
WHAT IT’S REALLY ABOUT
How Salesforce builds AI-native teams with ADLC, evals, and Slack
- Jayesh Govindarajan describes how Salesforce’s AI engineering org builds and operates predictive, generative, and agentic AI across the Salesforce stack, with Agentforce as a foundation customers use to build their own agents.
- He explains how AI-native work changes day-to-day execution: teams rely on tools like Claude/Codex for development, operate heavily in Slack, and increasingly manage “teams of agents” rather than doing all work by hand.
- He outlines an approach to adopting fast-moving AI tools: pass security requirements, give teams choice, observe real usage and adoption, then lean into emerging best practices while staying open to change.
- He details how to get determinism and “ground truth” from probabilistic AI by grounding agents in context and tool calls, using controls like Agent Script and adopting an eval-first culture with outcome-based testing and custom scoring.
- He introduces ADLC (Agent Development Life Cycle) as a new lifecycle for building, evaluating, deploying, and observing agents in production—where drift is expected and continuous iteration is necessary—plus explains how customer feedback and forward deployed teams shape roadmaps.
IDEAS WORTH REMEMBERING
5 ideasADLC replaces SDLC assumptions when the “software” is an agent.
He frames ADLC as the lifecycle for agentic software: build, test, evaluate, deploy, then observe so the agent can continually improve after launch, including responding to drift when context or underlying models change.
AI-native org structure mirrors the agent stack: models → harness → lifecycle/optimization.
He describes Salesforce’s AI engineering org as three stacked layers: teams that build/tune models, teams that build harness technologies (context, tool calling, memory), and teams focused on ADLC/agent optimization and production operations.
Roll out AI tools by enabling choice, then follow adoption data—not mandates.
He says tool choice should be open (after security review), with teams trying multiple options (e.g., Cursor, Claude Code, Codex) and leaders watching real usage data to see what drives adoption and best practices.
Headless + APIs lets customers consume Salesforce capabilities inside Claude or Slack.
He says moving “headless” is liberating because Salesforce’s “magic” is in APIs, metadata, and business logic—not just the UI—and bringing these capabilities into tools like Claude lets customers use Salesforce workflows in whatever interface they want.
Ground truth in probabilistic AI comes from tool-calling controls and structured agent specs.
He explains that grounding and determinism come from relying on deterministic APIs/tool calls rather than letting an LLM invent outputs; he describes Agent Script as a way to specify agents in plain English that transpiles into tool-calling logic and context retrieval, reducing hallucinations.
WORDS WORTH SAVING
5 quotesIf the software you're producing is not just a deterministic piece of software but is an agent, what does-- how do you build it? How do you test it? How do you evaluate it? How do you deploy it?
— Jayesh Govindarajan
I think we create them in our own heads, and, and, and then we specialize around them.
— Jayesh Govindarajan
We've had people change how they think about work, and that change is, in essence, how do you manage a team of agents, uh versus doing the work by hand, if that makes sense.
— Jayesh Govindarajan
The one place that people are available asynchronously all the time? It's Slack.
— Jayesh Govindarajan
The bottleneck is no longer writing code. The bottleneck is writing quality code, writing code that works all the time, having the test cases that fail before it hits production.
— Jayesh Govindarajan
High quality AI-generated summary created from speaker-labeled transcript.
