💃 Foundation First, Agents Second 🕺

Good morning, Salesforce Nerds! We’ve all seen them … the keynote demo was flawless. 🤌 

The agent resolved the case, updated the record, and charmed the customer in under thirty seconds. 🎤

Then you got back to your org. The one with three fields named Status__c, a Flow nobody will touch, and knowledge articles last updated when Aura was cool.

Here's the uncomfortable math: only about 7% of enterprises say their data is AI-ready. Meanwhile, roughly 95% of AI pilots fail. 📉

That’s a foundation problem, not a technology one.

And here's the good news buried in that stat: if most orgs aren't ready, readiness is the competitive edge. The teams doing foundation work now will ship agents that actually work while everyone else debugs hallucinations in production.

Let's build that foundation. 👇️

TABLE OF CONTENTS

💃 Foundation First, Agents Second 🕺

• Ground Control To Data Cloud

• Clean Metadata, Smarter Agents

• Trust Is An Architecture

• Observability Before Autonomy

• The Foundation Is The Feature

GARBAGE IN, AGENTS OUT

GROUND CONTROL TO DATA CLOUD

Agents don't know things. They retrieve things. 🧠

Every Agentforce response is grounded through retrieval-augmented generation: the Atlas Reasoning Engine pulls from CRM records, Knowledge, and Data Cloud, then reasons over what it finds.

Grounding quality is the single biggest determinant of agent reliability. Full stop. 💯 

Which means your duplicate contacts, half-empty opportunity records, and contradictory knowledge articles aren't cosmetic issues anymore. They're the agent's brain.

Data Cloud will happily unify a mess into a unified mess. 🗑️

The foundation work looks like this: deduplicate and complete your core objects. Audit Knowledge for accuracy and coverage. Ingest and index the unstructured content your agents will actually need, from PDFs to transcripts.

Then build retrievers deliberately. Scope what each agent can search, and verify retrieval quality at the prompt template level before an agent ever touches it. 🔍

Grounded agents see 25–40% better resolution rates than ungrounded ones.

That's not a tuning tweak. That's the whole ballgame. 📊

YOUR ORG IS THE PROMPT

CLEAN METADATA, SMARTER AGENTS

Here's the part nobody puts on a slide: your metadata is now part of the prompt. 🏗️

When an agent invokes an action, it reads your Flow labels, field names, descriptions, and picklist values to decide what to do. A field named Field_23__c with no description isn't just lazy.

It's misinformation you're feeding a reasoning engine. ❌ 

Agent actions are your existing automation wearing a new badge: Flows, invocable Apex, prompt templates, and MuleSoft APIs.

If that automation is brittle, undocumented, or tangled, you've just handed an LLM the keys to your technical debt. 🔑 

The foundation work: name things like a human will read them, because now a machine will. Write descriptions on every field, Flow, and action an agent can reach.

Refactor the automation you plan to expose. Small, single-purpose, transaction-safe actions beat sprawling mega-Flows every time. 🧹

Think of it as domain-driven design for agents. Each action should do one thing, declare its inputs clearly, and fail loudly.

An agent choosing between well-named, well-described actions is reasoning. An agent choosing between mystery Flows is gambling. 🎰

LEAST PRIVILEGE OR BUST

TRUST IS AN ARCHITECTURE

Agents operate within the permission model of the user or the agent user running them. They respect field-level security and can't touch data their context can't see. 🔐

That's a feature, but only if your security model deserves the trust.

Fifteen years of profile sprawl, "temporary" permission sets, and View All grants become agent-shaped attack surface the moment you go live.

An over-permissioned agent user is an incident report waiting for a timestamp.

The foundation work: create dedicated, least-privilege permission sets for each agent. Scope object, field, and action access to the job at hand. Review quarterly, because access always creeps. 🛡️

The Einstein Trust Layer handles the rest of the perimeter: PII masking before prompts reach the LLM, zero data retention with model providers, prompt injection defense, and a full audit trail of every interaction. ✅

But the Trust Layer sits on top of your security model. It doesn't replace it.

A masked prompt built from over-shared records is still a leak with extra steps. 🚫

WATCH YOUR ROBOTS WORK

OBSERVABILITY BEFORE AUTONOMY

Agents are probabilistic. They won't behave the same way twice, and that's by design. 🎲

Deterministic systems get unit tests. Probabilistic systems need evaluation harnesses, and you need them running before launch, not after the first executive escalation.

The foundation work starts in Testing Center: build scenario suites covering expected cases, edge cases, misuse attempts, and prompt injection. Define acceptance criteria per topic and action. 🧪

Then wire up production visibility. The Session Tracing Data Model logs every turn, reasoning step, LLM call, and guardrail check into Data 360, and it's OpenTelemetry-compliant, so traces flow into your existing observability stack. 📡

Finally, design your guardrails as architecture: strict grounding scope, topic boundaries that route out-of-scope requests to humans, and human-in-the-loop approval on high-stakes actions.

Retrofitting guardrails after launch costs multiples of designing them in. Ask anyone who's tried. 🚨

BUILD IT RIGHT FIRST

THE FOUNDATION IS THE FEATURE

Notice what the foundation actually is: clean data, honest metadata, least-privilege security, tested automation, real observability. 🏁

That's not an AI project. That's platform engineering you should have anyway. Agentforce just raised the stakes on skipping it.

The orgs winning with agents in 2026 aren't the ones who bought first. They're the ones whose platforms were ready when the agents arrived. 💪

Your competitors are demoing agents on quicksand. Build yours on bedrock.

The foundation isn't the boring part before the AI.

The foundation is the AI strategy. 🚀

SOUL FOOD

Today’s Principle

and now....Salesforce Memes