01 · Salesforce · Flow & Agentforce

Agentic Automation

Flow Powers Agentforce.

An Agentforce agent plans with a language model, so its reasoning is probabilistic. The work it does runs through Salesforce Flow and API-based actions, so the outcome is deterministic. My teams design the surfaces where admins and developers build those actions, and where people can see what an agent did.

Summary

The agent plans probabilistically. Salesforce Flow and API-based actions execute deterministically: same input, same outcome, every run
One action model: a flow, an @InvocableMethod, a prompt template or an API becomes an agent action, reused, never rebuilt
Agentforce passed $1B ARR and delivered 3.8B agentic work units in Q1 FY27, on a runtime running 461B+ Flow executions a month
Customers run 13 agents each, up from 5; actions per account compound 31% a month
My teams design the surfaces: Agentforce Builder, Agent Actions, Flow Builder, Testing Center and Flow Logging

youtube.com Custom agent actions, live

Apex and Flow become actions

youtube.com Agentforce Builder

Topics, actions, a live build from a blank slate

youtube.com Draft with Agentforce

Plain language to a working flow

engineering.salesforce.com 461B monthly executions

The runtime every action lands on

461B+

monthly Flow executions

^{Salesforce Engineering: Flow generation across 461B monthly executions, 2026}

$1B+

Agentforce ARR, Q1 FY27

^{Salesforce Q1 FY27 results, 27 May 2026}

3.8B

agentic work units in Q1, up 111% on the quarter

^{Salesforce Q1 FY27 results, 27 May 2026}

13

agents per customer, up from 5 in a year

^{Salesforce Agentic Enterprise Index, 2026}

31%

monthly compound growth in actions per account

^{Salesforce Agentic Enterprise Index, 2026}

99%

faster flow generation, 10 min to 10 s

^{Salesforce Engineering: Agentforce for Flow, 99% faster, March 2025}

01 · STAR

Situation

20%

The context a reader needs, and no more.

A customer asks the agent to move a delivery. A service rep asks it to escalate a case. Both need it to happen in Salesforce, correctly, every time, and the admin who owns the process needs to see what the agent did.

01

The customer and the rep

Want the outcome, not the conversation: a delivery moved, a case escalated, first time.

02

The admin and developer

Have years of flows and Apex encoding how the company works. Rebuilding that for agents was never an option.

02 · STAR

Task

10%

What I took responsibility for.

Make Flow and API-based actions the agent’s hands. The model decides what to do; the action does it, deterministically. One action model an admin can build, a developer can code, the reasoning engine can choose and a customer can trust.

03 · STAR

Action

60%

What I did, and for whom.

I organised the work around four people and one contract. The people: the admin, the developer, the agent and the customer. The contract: an action’s inputs, outputs and instructions, the whole interface between a probabilistic planner and a deterministic runtime.

01

One action, several doors

A flow, an Apex @InvocableMethod, a prompt template or an API becomes an agent action in Setup. The same action runs in Flow and in an agent, so nothing is built twice.

02

Instructions are the interface

The reasoning engine plans from labels, descriptions and instructions. The surfaces make people define the contract first and write instructions with example utterances.

03

Probabilistic plan, deterministic outcome

The model chooses the action. Flow or the API executes it: same input, same output, with the platform’s security on every run. Human approval is a flow element, so a step that needs a person waits for one.

04

Read, watch, repair

Agentforce Builder shows the plan and tests the conversation; Testing Center runs batches; Flow Logging records every run. Fix the flow and the agent picks it up.

05

The loop closes

Agentforce for Flow drafts a flow from plain language under a constrained DSL, 63+ flow types, and summarises a complex one in seconds. That flow is itself an action.

04 · STAR

Result

10%

What changed, with a number and a source.

Flow is the primary execution path for Agentforce actions.

01

Business

Agentforce passed $1B ARR; 3.8B agentic work units in a quarter, up 111%, on a runtime already at 461B+ Flow executions a month.

02

Product

One action model across Flow, Apex, prompt templates and APIs. 13 agents per customer, up from 5; actions per account compounding 31% a month; agent creation 53% faster; flow generation 99% faster.

Trust lives in the deterministic layer.Instructions are an interface.Reuse beats rebuild.