Agentforce
Agents that take real action — topics, custom Apex actions, grounding, the Einstein Trust Layer, and evals. Written while shipping a live, deployed agent.
Call It, Don't Hand Off
Transitioning to a connected subagent hands control away and the answer never comes back. Invoking it as an action returns the response to the router — the one difference that turns three specialists into a single reply.
Multi-Agent Orchestration: Who Owns What, and Who Answers
A router agent above three specialists — records, triage, scheduling — with exclusive ownership, a fixed order, and escalation held by exactly one of them. The design rules I only learned by publishing five versions of each.
An Honest Unknown Beats a Silent Pass
In a safety tool, the dangerous output is not the false alarm — it is the green light nobody investigates. Why my analyser reports 'we cannot determine this' as its own result, and why my evidence pack ships with two deliberate gaps.
The Engine Decides, the Agent Explains
My compliance agent answered the same legal question twice and gave two different wrong answers, both assembled out of its own instructions. The fix was to move the determination into a formula field and Apex — and to leave the agent no arithmetic to do.
Commit the Test Before the Grader
The order in which you commit an attack corpus and the harness that scores it is what makes the result believable. Pre-registration costs nothing, git does the notarising, and it is the difference between a result and a claim.
Never Let the Model Grade Itself
The most tempting shortcut in agent testing is to ask a language model whether the agent behaved. Here is why the grader has to be plain code — canary tokens, trace assertions, and what my own attack bench reported once it stopped being polite.
The Compliance Labels Your Org Already Has
Before you build a spreadsheet of regulated fields for your agent review, check the org: every custom field can carry a compliance classification, and code can read it back. That makes it a far better source of truth than a list you maintain by hand.
v67 Changed the Default, and Your Old Apex Did Not Notice
One org can hold two classes running the identical query with different effective access, purely because of the API version stamped on each file. What that means for an agent assembled from Apex written across several years.
The Escalation Gap: What Your Agent Reaches Beyond Its User
A permission set tells you what a person was granted. It cannot tell you what the code they trigger actually reaches. That difference is measurable — here is how I measure it, and how to do a rough version by hand.
Your Agent Runs As the User. That Is Not the Same As Seeing What the User Sees
Every Agentforce action runs as the authenticated user — a true sentence I had quietly turned into a promise about fields. Then I found a field in my own prompt template that the running user was never allowed to read.
A Green Dashboard Is Not a Healthy Agent
My runtime watchdog's heartbeat stayed green for a full day while every unit of work it enqueued died. What an uncatchable Apex error taught me about monitoring, Decimal arithmetic, and why absence is the only honest detector.
Red-Team Your Own Agent — Before Someone Else Does
I built a live Agentforce agent for a fictional DACH energy retailer (demo org) — then built the attack bench that tries to break it. Pre-registered attacks, a deterministic verifier and a cheap config lint, explained for beginners.
Grounded Actions: Designing a Multi-Turn Agentforce Flow That Actually Does Things
A walkthrough of a real, deployed Agentforce service agent: four grounded Apex actions, a multi-turn flow from 'who are you?' to a logged Case, and the habits — bulk-safe services, user-mode queries, mock-first testing — that kept it honest and affordable.
The Agentforce Planner: How Your Agent Decides What to Do
Topics, actions, and grounding are the parts — but something has to choose which action to run for each request. That something is the planner. Here's how an Agentforce agent actually decides, explained for beginners.
Fixing Agentforce 'Actions: 0' When Your Custom Apex Actions Don't Show Up
Your custom Apex agent action is deployed, but the Agentforce planner shows 'Actions: 0' and never calls it. Here's why the UI and metadata get out of sync — and the reliable fix.
Testing Your Agentforce Agent: Why and How
Agents are non-deterministic and customer-facing, so testing isn't optional — here's how to test, watch the planner, and refine until your agent is reliable.
Einstein, Copilot, Agentforce: Making Sense of Salesforce AI Names
Einstein, Copilot, Agentforce — the Salesforce AI names overlap and change, so here's a clear mental map of what each one actually means.
Building Your First Agentforce Action Without Writing Apex
You don't need to be a developer to give an Agentforce agent a real, working capability. Here's how to build your first action with Flow — the concept, the steps, and the beginner traps to avoid.
Prompt Templates in Salesforce: Reusable AI Instructions
Prompt templates turn one-off AI requests into reusable, grounded instructions — here's why they bring consistency and governance to your generative AI.
Grounding Explained: How to Stop Your AI Agent From Making Things Up
Grounding ties your Agentforce agent to real org data so it answers from facts instead of guessing — here's how it works and why it builds trust.
Topics and Instructions: Teaching Your Agent What It's For
Topics define what your Agentforce agent can help with, and instructions guide how it behaves — here's how clear scope keeps it reliable.
What Is Agentforce? A Plain-English Introduction for Salesforce Beginners
Agentforce is Salesforce's big bet on AI agents — but what is an 'agent,' really, and how is it different from a chatbot? A jargon-free introduction with the four ideas you need.