Integration Services
Business Consulting & Advisory
AI Solutions
AppExchange Development
Tailored Salesforce Engineering
Managed Services
Implementation
Dreamforce 2026 closed on September 17, and the Dreamforce 2026 announcements gave us a fair test. Before the event, we jot down what we expected to see: a shift from what agents can do to how you control them, a pricing story told carefully, customer proof worth questioning, and data as the part that decides whether any of it works.
This Dreamforce 2026 recap checks those calls against what Salesforce shipped, put into pilot, or dated for later. It is for the people who now have to turn a keynote into a budget line and a release plan.
The short version: the direction held. What we underestimated was how far Salesforce would go in treating its own screen as one surface among several.
Read the Dreamforce 2026 Announcements as One Stack
Salesforce’s own recap lists 211 announcements. Read as layers, they make sense: where work happens (AIforce), who does it (agents and models), and how you watch it (testing and observability).
Watch the status column. A pilot does not belong in a Q1 release plan.
| Release | What it does | Status per Salesforce, Sept 28 |
|---|---|---|
|
AIforce: Claudeforce, Slackforce, Agentforce Coworker |
Brings Salesforce data, workflows, and your existing permissions into Claude, Slack, and Lightning |
Coworker GA; Slackforce GA October; Claudeforce beta |
|
Job-ready agents (Piper, Fin, Hunter and others) |
Prebuilt agents for named roles such as inbound SDR and customer service |
Most GA; Hunter GA November |
|
Marshall and Systems Integration Agent |
Builds a fixed, approved path for back-office processes; works inside browser-based apps without APIs or RPA |
GA |
|
Long-horizon agents |
Pursue one goal over days or weeks, with memory across sessions and approval checkpoints you define |
Pilot; GA November 2026 |
|
Koa |
Salesforce’s first CRM reasoning model, built on NVIDIA Nemotron and trained on synthetic data |
Pilot; GA expected Winter 2026 |
|
Agent Optimizer |
Reads production sessions, ranks recurring failures by impact, and stages fixes |
GA October |
|
Agent Health Monitoring, Custom Scorers, A/B Experimentation |
Production alerts across 16+ metrics, grading logic you define, and variant testing |
Monitoring GA; Scorers API GA, UI beta; A/B beta October |
|
Einstein Bot upgrade |
Converts an existing Einstein Bot into an Agentforce agent while the bot keeps running |
Available through Agentforce Labs |
Govern Agents Like Production Code: The Control Call Held
We said the question would move from whether an agent can do something to whether you can predict and prove what it did. That held. Most of the releases that matter for production are control features.
- MCP Risk Scores. Agentforce now scans an MCP server at registration2 for prompt injection, tool poisoning, and rug pull attacks, rates it Low, Medium, or High, and lists fixes before the connection is approved. Scanning continues after registration.
- Fixed paths for back-office work. Marshall builds an approved path for each process, so the same inputs produce the same results. That is the deterministic logic we flagged, now packaged as an agent.
- A named architecture. Salesforce grouped all of this under an Enterprise AI Harness3 with six layers: models, context, agency, actions, governance, and security, run from one AI control plane.
What we missed was AIforce. With Claudeforce and Slackforce, a CRM action can now start in Claude or Slack, governed by the permissions you already have. That makes your permission model the most important control you own. If profiles and permission sets have drifted, every new surface inherits the drift.
Switching on Salckforce or Claudforce? Get a Clear Read of Who can See and Change What.
Price Agent Work by the Credit: The Model Arrived Before the Keynote
We expected Dreamforce to lean on consumption and credit pricing. Salesforce moved first. On September 3, it replaced its edition4 lineup for new buyers of Agentforce Sales, Agentforce Service, and Agentforce Industries.
| Edition | List price (USD per user/month) | Flex Credits included |
|---|---|---|
|
Core |
$195 |
500,000 |
|
Advanced |
$395 |
1 million |
|
Max |
$550 |
2.75 million |
Each edition bundles Agentforce, Slack, Tableau Next, data security, and the Premier Success Plan. Existing customers keep legacy pricing, and Agentforce 1 customers can move to Max at no extra cost.
A standard agent action costs 20 Flex Credits, or $0.10, per Salesforce’s Agentforce pricing page5. That puts Core’s allocation at 25,000 standard actions a year for the whole org, not per seat. And unused Flex Credits do not roll over into the next term.
The outcome-based signal we said to watch for did appear, but narrowly. The Help Agent in Advanced and Max is priced on resolutions, listed at $2 each. Everything else still meters activity.
Because these editions are bundles, our pre-event advice applies directly. Before your renewal, list what you use, line by line, and model expected monthly actions against the included credits. If you want that math checked before you sign, talk to our Salesforce consulting team.
Judge Agents on Outcomes: Proof Became Tooling
The keynote stories came from a direction different from our prediction. Fulton Bank6 went from first hearing about Agentforce to a working agent in an hour, and to 3,000 employees about seven weeks later. Southwest Airlines reported productivity gains and an ROI multiple, without the cost base behind it.
Both are adoption stories. Neither, as told on stage, answered the questions we said we would listen for: how many people maintain the agents, what broke in month three, and what the company stopped doing.
Salesforce now ships tools to answer them yourself. Agent Optimizer ranks recurring failures from production sessions. Voice and Conversation Testing scores whether an agent reached the goal, not just whether each turn was correct. Custom Scorers let you define what a good outcome is. Treat these as the evidence layer for your next budget review.
Fix the Data Before You Widen the Surface
Salesforce wrote data into the harness, where governance covers data quality and classification. AIforce widens the exposure. A customer record that disagrees across three systems now shows up in Slack and Claude as well as Lightning. See how we approach Data 360 readiness before agents reach that data.
Turn Dreamforce 2026 Announcements into a 90-Day Plan
The useful part of any Dreamforce 2026 recap is the work it creates. Between now and Spring ’27, that work is mostly readiness.
- 1. Plan production work only on GA items, such as Coworker and Agent Health Monitoring. Treat Koa and long-horizon agents as pilots.
- 2. Review profiles, permission sets, and sharing rules before any AIforce surface goes live.
- 3. List every MCP server and external tool your agents reach, and register them so Risk Scores run.
- 4. Model credit use per agent before your renewal conversation, not during it.
- 5. If you run Einstein Bots, test the upgrade path in a sandbox while the bot stays live.
- 6. Pick one production agent and define a scorer for the outcome you care about.
Two decades of Salesforce implementations have taught us that the projects that go well start with an honest look at the org you already have. As a Salesforce Summit Partner, that is where we start too.
Sources :
