Dreamforce 2026 put one idea at the center of Salesforce’s product strategy: enterprise AI is moving from answering questions and completing isolated tasks toward executing longer-running business processes.
Held in San Francisco from September 15–17, the event brought a large wave of announcements across Salesforce, Agentforce, Data 360, Slack, AI infrastructure, developer tools, and enterprise governance.
Salesforce’s post-event roundup lists 21 major announcements, ranging from its new CRM reasoning model, Koa, to AIforce, long-horizon agents, reusable AI Skills, Agent Optimizer, job-ready agents, and new ways of bringing Salesforce capabilities into other interfaces.
Rather than covering every announcement individually, this guide focuses on the developments that could have the biggest technical and architectural implications.
Table of Contents
- AIforce: Salesforce Beyond the Traditional UI
- Agentforce Coworker: An AI Teammate Inside Salesforce
- Koa: Salesforce’s CRM Reasoning Model
- Long-Horizon Agents
- Agent Optimizer
- AI Skills
- Job-Ready Agents
- Marshall and Back-Office Automation
- Data 360 and Trusted Enterprise Context
- Slack as an AI Workspace
- Headless Salesforce and MCP
- The Enterprise AI Harness
- What Dreamforce 2026 Means for Salesforce Developers
- What It Means for Salesforce Admins and Data Teams
- Frequently Asked Questions
- Conclusion
AIforce: Salesforce Beyond the Traditional UI
One of the biggest announcements at Dreamforce 2026 was AIforce, Salesforce’s new interface layer designed to bring Salesforce data, workflows, business logic, permissions, security, and governance into different AI interfaces.
The idea represents a significant shift from traditional enterprise software.
Historically, employees opened an application, navigated through menus, located a record, performed an action, and then moved to another application.
Salesforce is now pushing toward a model where the user can interact with business capabilities wherever they already work.
AIforce includes experiences such as Claudeforce inside Claude, Slackforce inside Slack, and Agentforce Coworker inside Salesforce. Salesforce says these experiences can use existing Salesforce permissions and governance rather than creating an entirely separate security model.
The architectural idea can be represented as:
Salesforce
│
┌─────────────┼─────────────┐
│ │ │
Data Workflows Business Logic
│ │ │
└─────────────┼─────────────┘
│
AIforce Layer
│
┌────────────┼────────────┐
│ │ │
Slack Claude Salesforce
Instead of forcing employees to move into Salesforce for every task, Salesforce is attempting to make its capabilities available across the places where work happens.
That could have significant implications for enterprise application architecture.
Agentforce Coworker: An AI Teammate Inside Salesforce
Another major announcement was Agentforce Coworker, which Salesforce describes as an autonomous AI teammate built directly into Salesforce.
Users can ask it questions in natural language, and it can search across CRM data, Slack, and other connected sources. It can also create plans and execute tasks, including orchestrating other agents.
This is different from a conventional chatbot.
A chatbot generally follows a conversational pattern:
Question
↓
Retrieve information
↓
Generate answer
An agentic workflow looks more like:
Goal
↓
Understand context
↓
Create plan
↓
Select tools
↓
Execute actions
↓
Evaluate result
↓
Continue / adjust
The distinction matters because enterprise agents increasingly need to operate across multiple systems rather than simply retrieve information.
Koa: Salesforce’s CRM Reasoning Model
Salesforce also introduced Koa, its first CRM reasoning model for Agentforce.
According to Salesforce, Koa is built on NVIDIA Nemotron and designed specifically for CRM-oriented reasoning, tool use, and decision-making. Salesforce says the model’s training corpus uses synthetic scenarios rather than real customer data and that the company operates the model within its own infrastructure.
The broader strategy is worth watching.
Instead of depending exclusively on general-purpose foundation models, Salesforce is developing a model optimized for its own business domain.
The reasoning is similar to specialized models in other technical domains:
General-purpose model
↓
Broad capabilities
CRM-specific model
↓
CRM terminology
CRM workflows
CRM actions
CRM reasoning
Salesforce reported internal pilot results indicating improvements in context retention, answer relevance, and tool-call precision compared with its internal benchmarks. Those figures are Salesforce’s own reported results, rather than independent benchmark measurements.
For developers, the more important question may be how specialized reasoning models eventually interact with external models.
Enterprises could increasingly choose different models for different tasks based on cost, latency, reasoning requirements, security, and domain specificity.
Long-Horizon Agents
One of the more technically interesting announcements was Salesforce’s introduction of long-horizon agents.
Traditional AI agents often perform relatively short tasks:
Find the customer record.
Summarize this opportunity.
Draft an email.
Long-horizon agents are designed to pursue goals over much longer periods—potentially days, weeks, or months.
Salesforce gives the example of re-engaging at-risk deals.
Instead of completing one prompt, the agent can develop a multi-step plan, execute actions over time, request approval when required, and adjust its plan when circumstances change.
Salesforce says the new runtime provides three important capabilities:
- Memory to retain context across sessions
- Durable execution to recover when steps fail
- Dynamic steering to allow users to adjust plans conversationally
The architecture becomes closer to:
Business Goal
↓
Planning
↓
Task 1 ──► Task 2 ──► Task 3
│
▼
New information
│
▼
Re-plan / Adjust
│
▼
Continue goal
Salesforce listed long-horizon agents as a pilot at Dreamforce, with general availability planned for November 2026.
This is an important development because persistent agents introduce new requirements around monitoring, permissions, state management, approvals, and failure recovery.
Agent Optimizer
Salesforce also announced Agent Optimizer, designed to help build, test, and improve other agents.
According to Salesforce, Optimizer can help create subagents and actions, test them during development, and analyze production sessions to identify recurring failure patterns.
This creates an interesting feedback loop:
Agent
↓
Production Sessions
↓
Observability
↓
Failure Patterns
↓
Agent Optimizer
↓
New Version
↓
Testing
↓
Deployment
This is effectively bringing software-development and observability concepts into agent development.
Instead of treating an AI agent as a static configuration, organizations can begin thinking about an agent lifecycle:
Build → Test → Deploy → Observe → Improve → Govern
Salesforce announced general availability for Agent Optimizer in October 2026.
AI Skills: Build Once, Reuse Across Agents
Another announcement was AI Skills.
Salesforce describes AI Skills as reusable, governed instructions that can be used by multiple agents.
For example, an organization could define a standardized skill for:
- Qualifying a lead
- Checking account status
- Preparing a customer summary
- Escalating a support issue
- Performing a compliance check
Instead of rebuilding the same logic into multiple agents, the organization can create the capability once and reuse it.
AI Skill
│
┌───────────┼───────────┐
↓ ↓ ↓
Agent A Agent B Agent C
This could become particularly important as companies deploy larger numbers of agents.
Without reusable capabilities, every new agent can introduce duplicated instructions, inconsistent behavior, and additional maintenance.
Salesforce listed AI Skills as a pilot with general availability planned for October 2026.
Job-Ready Agents
Salesforce is also moving beyond the idea that every organization must build its own agents from scratch.
At Dreamforce, the company highlighted job-ready agents designed for specific roles and workflows.
Examples include:
- Piper for sales development
- Fin for customer service
- Hunter for sales
- Marshall for supply-chain and back-office processes
Salesforce says these agents come with skills and actions designed for their specific roles and can be continuously taught as business requirements evolve.
This reflects a broader change in enterprise AI.
The market is moving from:
“Here is a model. Build something.”
toward:
“Here is an AI worker designed for a particular business process.”
The technical challenge then shifts from model development toward configuration, integration, governance, monitoring, and process design.
Marshall and Back-Office Automation
Marshall is particularly interesting because it focuses on back-office and supply-chain processes.
Salesforce says users can describe a process, answer clarifying questions, and allow the system to create the required steps, forms, and connections to systems of record.
The resulting workflow can then be executed through a Systems Integration Agent. Salesforce says it can work with browser-based applications without requiring traditional code, APIs, or RPA for those integrations.
If this approach works reliably at scale, it could affect how organizations approach legacy application integration.
However, enterprises will still need to consider security, auditability, application changes, error handling, and governance before allowing autonomous agents to operate against production systems.
Data 360 and Trusted Enterprise Context
AI models are only one part of an enterprise AI architecture.
The other major component is data and context.
Salesforce continues to position Data 360 as a foundation for trusted enterprise data and Agentforce.
At Dreamforce, Salesforce highlighted its Agentic Customer Data Platform and Agent Context Engine as part of its broader data strategy.
The architecture increasingly looks like:
Enterprise Systems
↓
Data Integration
↓
Data 360
↓
Trusted Context
↓
Agentforce
↓
Business Actions
This is important because an agent that has strong reasoning but poor business context can still produce unreliable outcomes.
Enterprise AI therefore requires more than an LLM.
It requires:
- Trusted data
- Metadata
- Permissions
- Business rules
- Context
- Integration
- Governance
- Monitoring
Slack as an AI Workspace
Slack also received significant attention at Dreamforce.
Salesforce increasingly positions Slack as a place where humans and AI agents can collaborate rather than simply as a messaging application.
The company also highlighted Slack Code, which allows developers to work with coding agents and create interactive experiences within conversations.
This supports a broader Salesforce strategy:
Bring AI to the user’s workflow instead of requiring users to constantly switch applications.
The implications extend beyond Slack.
Salesforce has been emphasizing a headless approach in which CRM intelligence, business logic, data, and agents can be surfaced through different interfaces.
Headless Salesforce and MCP
Another important technical direction is headless architecture.
Traditional Salesforce usage generally assumes that users interact through Salesforce interfaces.
A headless architecture separates the underlying business capabilities from the user interface.
Salesforce Platform
│
┌────────────┼────────────┐
│ │ │
API MCP Agents
│ │ │
└────────────┼────────────┘
│
Multiple User Interfaces
This allows Salesforce capabilities to appear in applications and AI interfaces outside the traditional Salesforce UI.
Salesforce has also emphasized MCP security and headless development as part of its IT strategy at Dreamforce.
For developers, this means Salesforce architecture increasingly needs to account for APIs, agent interfaces, permissions, identity, and machine-to-machine interactions, not just user-facing applications.
The Enterprise AI Harness
Perhaps the most important architectural idea underneath all these announcements is the AI harness.
Salesforce argues that powerful AI models alone are insufficient for enterprise deployment.
Organizations need controls around them.
Salesforce describes six capabilities for an enterprise AI harness, including trusted models, trusted context, trusted agency, trusted actions, governance, and security.
Conceptually:
AI Models
│
┌────────┴────────┐
│ AI Harness │
│ │
│ Context │
│ Permissions │
│ Actions │
│ Governance │
│ Security │
│ Observability │
└────────┬────────┘
│
AI Agents
This is arguably one of the most important lessons from Dreamforce 2026.
As AI becomes capable of taking actions—not just generating text—the surrounding control layer becomes increasingly important.
What Dreamforce 2026 Means for Salesforce Developers
For developers, the announcements point toward a Salesforce architecture that is becoming increasingly agent-oriented.
Traditional development skills remain relevant, but developers may increasingly work with:
- APIs
- MCP
- Agentforce
- AI Skills
- Agent orchestration
- Event-driven architectures
- Data 360
- Agent observability
- AI security
- Model selection
- Headless applications
The developer’s role may shift from building every workflow manually toward designing the systems, tools, APIs, data structures, guardrails, and integrations that agents use.
What It Means for Salesforce Admins and Data Teams
Admins will also have an increasingly important role.
More agents mean more questions around:
- Which agents can access which records?
- Which actions require approval?
- What data can an agent retrieve?
- How are agent actions audited?
- How are errors detected?
- Which skills can agents execute?
- How should agent performance be measured?
Data teams face a related challenge.
If enterprise AI depends on trusted context, then data quality, governance, metadata, identity resolution, and lineage become increasingly important.
AI does not eliminate the need for good data architecture.
It makes it more visible.
Frequently Asked Questions
Some of the major announcements included AIforce, Agentforce Coworker, Koa, long-horizon agents, Agent Optimizer, AI Skills, job-ready agents, and expanded headless and AI capabilities. Salesforce published a roundup covering 21 announcements from the event.
Koa is Salesforce’s CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Salesforce says it is designed for CRM-specific reasoning, tool use, and decision-making and was trained using synthetic scenarios rather than real customer data.
AIforce is Salesforce’s interface layer for bringing Salesforce data, workflows, business logic, security, permissions, and governance into different AI interfaces and work environments. Salesforce highlighted experiences including Claudeforce, Slackforce, and Agentforce Coworker.
Long-horizon agents are designed to pursue goals over extended periods rather than completing a single short task. Salesforce says its long-horizon runtime provides memory, durable execution, and dynamic steering.
Agent Optimizer is designed to help create, test, analyze, and improve Agentforce agents. Salesforce says it can analyze production sessions to identify recurring failure patterns and help stage improvements.
