My Technology Learnings (Salesforce): What's New from Dreamforce 2026

A layer that lets any AI tool or app use Salesforce data, workflows and actions through APIs and MCP, with the user’s security carried along on every call. It was introduced at TDX 2026 as Headless 360, launched at Dreamforce as AIforce, and renamed SIforce by Marc Benioff in October. The engine underneath is the Headless Toolkit.

Use it when your users live in Claude, Slack or your own app and shouldn’t have to switch to a Salesforce tab. It has three surfaces out of the box: Claude (Claudeforce), Slack (Slackforce) and Lightning (Agentforce Coworker).

Every AI tool needed its own custom integration, often running as an all-powerful integration user. Enforcing each person’s access was hard, and auditing it was even harder.

Claudeforce is the Salesforce and Anthropic partnership. Its first product, Salesforce in Claude, is a plugin built by both companies that brings live CRM data and 37 ready-made sales skills into Claude. It connects through Salesforce-hosted MCP servers via SIforce.

For sellers and managers who already plan their day in Claude. An admin connects the plugin once. Each user signs in to Salesforce with their own login (OAuth), so their own permissions decide what Claude can read or change.

You copied data into a chat window or built your own connector. Per-user security was hard to guarantee, and nothing could be written back safely.

Slackforce is the Slack part of SIforce. Slackbot, powered by Claude, reads and acts on Salesforce, Data 360 and Tableau data within your permissions, and it now also sits in Lightning as a chat panel next to the record. Slackforce Surfaces are dashboards, reports, decks or microsites that Slackbot builds from a plain-language request.

For teams that run their day in Slack. You can pin a Surface to a channel, where the team can filter it, comment on it and act on it together. Each person sees only what they are allowed to see, and live data refresh started rolling out in October.

Salesforce for Slack gave you record previews and alerts. Real dashboards meant leaving Slack, and building them took an admin’s time.

A ready-to-use AI teammate in the Lightning search bar. It is also available in Microsoft Teams, on mobile, and in Claude and ChatGPT via MCP. It answers questions, explains why something happened, builds a plan, calls your existing agents and flows, and turns repeated work into reusable Skills.

For every employee, without building a single agent. It is grounded in Salesforce data and enterprise data through Data 360, works within the signed-in user’s permissions, and an admin can set it up in minutes.

Each assistant had to be designed, configured and tested one by one as a task-specific agent, with its own topics, actions and tests.

Salesforce’s first CRM reasoning model. It is built on NVIDIA Nemotron 3 Super and post-trained by Salesforce on a synthetic CRM dataset covering 14+ industries, with no customer data. Salesforce owns the model weights and runs it inside its own trust boundary.

Choose it for agents that do multi-step CRM work: picking the right tool, in the right order, with the right inputs. You can select it for any agent, subagent or agent router in Agentforce Builder, or set it as the model for the whole org.

General-purpose models are excellent at language but aren’t trained on CRM workflows, so tool-calling mistakes were more common. Salesforce reports 3× fewer errors on CRM actions with Koa (on its own benchmark).

A new no-code workspace where you describe what you need in plain language. It generates the app, agent or experience, then tests and deploys it, with safety checks, automated testing and deployment controls built in. It is built on SIforce, so it respects your data, metadata and security.

For admins turning a business request into a working solution quickly. Developers who want code and source control have Agentforce Vibes, while Builder Central is the admin’s equivalent.

Admins assembled each piece separately in Object Manager, Flow Builder and Lightning App Builder, then tested and deployed by hand.

Observability (in Agentforce Studio) records every agent session in Data 360: user inputs, reasoning steps, actions, LLM calls and guardrail checks. It adds dashboards, health monitoring and, new this year, voice observability. Agent Optimizer reads those traces, ranks failure patterns and proposes configuration fixes along with regression tests. The Agent Router is the start_agent block in Agent Script, the default entry point that works out what the user wants and passes them to the right subagent.

Use these to run agents in production with confidence: see what happened, understand why, fix it and prove the fix. The router is the default, but it isn’t mandatory. You can make another subagent the starting point, and you can set Koa as the router’s model.

Debugging meant reading raw logs one conversation at a time. Improvements were guesswork, and regressions slipped through unnoticed.

A native security capability delivered through Security Center. It brings in security data from Salesforce and partner tools, normalises it into one schema (OCSF, the Open Cybersecurity Schema Framework) in Data 360, and correlates the streams to detect threats that cross systems.

For security teams protecting Salesforce alongside identity and endpoint tools. Sources available today are Real-Time Event Monitoring, Okta (identity security), CrowdStrike (endpoint security), DigitSec (code scanning) and WithSecure (file threat detection).

Security Center gave security-posture views across orgs (health checks, permissions, configuration drift), and activity logs lived in Event Monitoring. Linking those to external tools meant exporting everything to a separate security platform.

A control plane for every agent, API, MCP server and model call across the company, whichever platform built it. It has seven parts: Scanners, Agent Registry, Agent Broker, Omni Gateway (formerly Flex Gateway), Agent Visualizer, Governance and Monitor.

For architects facing agent sprawl: AI agents piling up across teams and clouds with nobody keeping track of what exists, what it can reach or what it costs. Scanners crawl Agentforce, Amazon Bedrock, Google Vertex AI, Microsoft Copilot Studio and others, on a schedule or on demand, and record what they find in the Registry as A2A agent cards.

Each platform kept its own list of agents. Flex Gateway governed APIs, but agent, MCP and LLM traffic had no single place for policy or visibility.