Salesforce Agentforce Job-Ready Agents and Long-Horizon Runtime: The Complete Guide for Agentic AI Architects (2026)

The Shift from “Build Your Own” to “Deploy and Customize”

On September 11, 2026 — four days before Dreamforce opens in San Francisco — Salesforce dropped its largest Agentforce announcement since the platform’s original launch. Seven named, job-ready AI agents. A long-horizon runtime that lets agents pursue goals across weeks, not conversations. A six-pillar Trusted Enterprise AI Harness that governs agents from any vendor. And a deepened Anthropic partnership, Claudeforce, that makes Claude the default reasoning model across Agentforce surfaces.

For agentic AI architects evaluating enterprise agent platforms, this isn’t incremental. It’s a structural shift: Salesforce is moving from selling an agent-building toolkit to shipping production-grade agents that arrive knowing how to do specific jobs — and a governance layer designed to manage the multi-vendor agent sprawl that 85% of enterprises are already living with, according to a July 2026 VentureBeat Intelligence survey.

This guide breaks down every component of the announcement, maps the architecture, evaluates what’s genuinely new versus repackaged, and tells you exactly where each piece fits in an enterprise agentic AI strategy.

The Seven Job-Ready Agents: Who They Are and What They Do

The centerpiece of the September 11 announcement is a portfolio of seven named agents, each purpose-built for a specific business function. Unlike the previous Agentforce model — where companies assembled agents from skills, actions, and prompts inside Agent Builder — these agents ship with pre-built data models, integrations, and domain-specific reasoning already wired in.

Here’s the complete roster:

Six of the seven are generally available today. Hunter — the outbound sales agent running on the new long-horizon runtime — is in pilot with GA targeted for November 2026.

Why Named Agents Matter Architecturally

The naming isn’t just branding. Each named agent represents a pre-configured agent template with its own skill graph, permission model, and data access patterns already mapped to Customer 360. When an architect deploys Casey for customer service, they’re not starting from a blank Agent Builder canvas — they’re customizing a working agent that already understands return policies, account lookups, and escalation trees within the Salesforce data model.

This is a direct response to the deployment gap that has plagued enterprise agentic AI. As Futurum Group’s analysis noted, the biggest barrier to agentic AI adoption isn’t the technology — it’s the weeks or months of custom development required to get an agent from “hello world” to production-ready. Salesforce is compressing that timeline by shipping the 80% that’s common across deployments and letting architects customize the 20% that’s unique to their business.

More Read

16 Reasons Why Agentic Automation Programs Fail – And How to Never Repeat Them

Microsoft Copilot Autopilot: The Always-On Enterprise AI Agent Guide for Agentic AI Architects

DigitalOcean Managed Agents and Agent Droplets: The Complete Guide for Agentic AI Architects (2026)

Customers can rename any agent to match their brand — Casey becomes “Eva” at Engine, Hunter becomes whatever fits the sales team’s culture — so the agent becomes an extension of the workforce rather than an obviously third-party tool.

Early Production Metrics

Salesforce published production metrics from early adopters alongside the announcement. These numbers come from live deployments, not demos:

• Engine: 50% of chat inquiries fully resolved by its help agent (Casey, renamed Eva)
• Perk: 60% of sales pipeline built by its outbound sales agent (Hunter)
• Autism Queensland: 70% of administrative requests resolved by its employee service agent (Paige)
• Hibbett: 90% of core shopper journeys handled by Hibbett AI, which went live in six weeks
• Asana: 4x conversation volume driven by its website agent (Piper), with average deployment time of 45 days
• Anthropic: 79% of conversations that Fin sees are resolved autonomously

The Hibbett figure is architecturally significant: 90% coverage in six weeks suggests the pre-built agent template dramatically reduced the integration and testing overhead that typically extends enterprise agent deployments to 3-6 months.

The Long-Horizon Runtime: Agents That Pursue Goals Over Weeks

The most technically significant part of the announcement isn’t the named agents — it’s the infrastructure underneath Hunter. Salesforce built a long-horizon runtime that enables agents to pursue goals across days and weeks, not just complete a task within a single conversation.

This is a fundamental architectural departure from how most enterprise agent platforms work today. The standard pattern — even in sophisticated frameworks like the A2A and MCP ecosystem — is request-response: a user sends an instruction, the agent completes a task (or chain of tasks), and returns a result. State ends when the session ends.

The long-horizon runtime changes this by adding three capabilities:

1. Persistent Memory

Memory preserves context and progress across sessions so work doesn’t stop when an interaction ends. When Hunter researches an account on Monday, that research is available on Thursday when the follow-up email needs to reference what was learned. This isn’t conversation history — it’s structured, goal-oriented memory that tracks what’s been tried, what worked, and what the current state of each sub-task is.

2. Durable Execution

Plans keep running over time and can resume or course-correct as circumstances change. If Hunter sends an outreach email and the prospect replies three days later with a question, the agent picks up the thread within the context of the original goal — not as a new, context-free interaction. The execution engine handles interruptions, delays, and external events without losing the thread.

3. Dynamic Steering

The agent adapts its behavior based on individual user feedback and direction. A seller can tell Hunter to change approach, reprioritize deals, or shift strategy mid-flight, and the agent adjusts the plan without starting over.

How It Works in Practice

Salesforce’s reference scenario illustrates the architecture: a seller asks Hunter to “rescue at-risk deals before end of quarter.” Hunter doesn’t just draft a list. It:

1. Converts the objective into a measurable goal with success criteria
2. Builds a multi-step plan spanning weeks
3. Determines the tasks, tools, and context required at each step
4. Applies guardrails defining when it can act autonomously and when seller approval is required
5. Executes the plan over time, incorporating new information and adjusting as the situation evolves

This is closer to how IBM watsonx Orchestrate’s control-plane architecture handles multi-step workflows than to a traditional chatbot loop. The difference is that Salesforce is building it natively into the CRM runtime rather than as a separate orchestration layer.

What This Means for Agentic AI Architecture

The long-horizon runtime is currently exclusive to Hunter, but Salesforce confirmed that more agents will run on it over time and that customers will be able to build their own long-horizon agents with Agentforce. For architects planning enterprise agent deployments, this creates a new design pattern: goal-oriented agents that maintain state across sessions, rather than task-oriented agents that complete discrete operations.

The implications are significant for how enterprises think about agent orchestration. Instead of building complex workflow chains where each step triggers the next agent, a single long-horizon agent can own an entire business objective and manage the sub-tasks internally. This reduces the orchestration overhead but requires more sophisticated memory management, guardrail design, and human-in-the-loop patterns.

The Trusted Enterprise AI Harness: Governing Multi-Vendor Agent Sprawl

Separately from the job-ready agents, Salesforce announced the Trusted Enterprise AI Harness on September 10 — a six-pillar governance framework for enterprises that are already running agents from multiple vendors.

The timing and positioning are deliberate. The VentureBeat Intelligence survey cited above found that 85% of enterprises run two or more agent orchestration platforms, with an average of 3.1 platforms per company. Salesforce is betting that even companies deeply invested in competing platforms (AWS Bedrock, Azure AI Foundry, Google’s Gemini Enterprise Agent Platform) need a unified governance layer — and that Salesforce’s position as the system of record for customer data makes it the natural home for that layer.

The Six Pillars

The AI Control Plane

The Harness includes a new AI Control Plane — a management layer that sits above any single agent platform. According to Salesforce’s official announcement, the capabilities live in a shared harness rather than inside individual agents, making them reusable across the enterprise. Context can be shared across agents and models, and governance and security are applied consistently regardless of which vendor built the agent.

The architectural decision to build the Harness headlessly is the most important detail for architects. The capabilities are accessible through MCP (Model Context Protocol), APIs, Skills, and Plugins — meaning they extend beyond traditional Salesforce applications into AI experiences where work already happens. Salesforce explicitly listed Claude, Slack, Microsoft Teams, and other AI experiences as target surfaces.

This positions the Enterprise AI Harness as a competitor to IBM watsonx Orchestrate and ServiceNow AI Control Tower in the emerging agent governance space — but with the advantage of sitting on top of Customer 360 data that most enterprise agents eventually need to access anyway.

Claudeforce: The Anthropic Partnership Deepens

On August 26, 2026, Salesforce and Anthropic announced Claudeforce — a strategic partnership that goes well beyond the typical “we integrated their model” announcement. The details, published in a joint press release, reveal a deep operational integration:

• Claude as the default model: Claude is now the default reasoning model for Slack AI, Slackbot, Salesforce in Claude, Headless 360, and Agentforce Coworker
• Salesforce plugin inside Claude: A plugin with 37 prebuilt sales skills — meeting prep, deal health review, pipeline review — lets sellers and agents reason over live CRM data directly from Claude’s interface
• Claude as the first LLM fully inside the Salesforce Trust Boundary: This means Claude has access to Customer 360 data under the same security and compliance controls as native Salesforce features
• Claude Code and Claude Enterprise for all Salesforce developers and knowledge workers: Salesforce is making these Anthropic’s preferred AI assistant and productivity tools across its own engineering organization

For agentic AI architects, the Claudeforce partnership creates a unique architecture pattern: agents built on Agentforce can use Claude for reasoning while Agentforce handles the CRM data layer, business rules, and governance. And separately, Claude users can access Salesforce data and actions through the plugin without leaving Claude’s interface.

Salesforce in Claude is available to select pilot customers now, with open beta expected in September 2026 and additional prebuilt skills launching in late 2026.

New Editions: Core, Advanced, and Max

On September 3, Salesforce restructured its pricing into three editions — Core, Advanced, and Max — for Agentforce Sales, Agentforce Service, and Agentforce Industries. Each edition bundles Slack, Tableau Next, Premier Success Plan, enterprise-grade security, and AI capabilities into a single SKU.

The pricing consolidation matters for architects building business cases. Instead of assembling a stack of individual SKUs (Salesforce license + Einstein AI + Slack + Tableau + Security add-ons), the editions bundle everything needed for an agentic deployment. Existing Agentforce 1 Edition customers can upgrade to Max at no additional cost.

The Flex Credit model — where agent actions consume credits rather than requiring per-agent licensing — aligns with the consumption-based pricing that most agent platforms are converging on. Salesforce’s Agentic Work Unit (AWU) metric, which measures discrete tasks accomplished by AI agents, provides the measurement layer: the company reported 7 billion AWUs delivered across Agentforce and Slack, with 3.2 billion in Q2 2026 alone, growing at a 15% compound monthly growth rate.

The Agentforce Platform: Multi-Agent Orchestration and Continuous Improvement

Beyond the named agents, Salesforce expanded the underlying Agentforce platform with three capabilities that change how architects design multi-agent systems:

AI Skills in Agentforce Coworker

Employees can teach Agentforce Coworker how to complete a task once, then scale that knowledge across the workforce and across interfaces. This is essentially a no-code skill-creation mechanism: a subject matter expert demonstrates a workflow, and the platform converts it into a reusable agent skill. Pilot now, GA October 2026.

Multi-Agent Orchestration

Routes work across agents so they operate as one coordinated team when a job crosses roles, systems, or stages of the customer journey. This is GA now and represents Salesforce’s answer to the Agent-to-Agent (A2A) protocol and similar orchestration standards — though built natively within the Salesforce ecosystem rather than as an open protocol.

Agent Optimizer

An agent that helps teams build, refine, test, and analyze other agents. It works across the agent lifecycle — from initial construction through production monitoring — analyzing session traces to identify what to improve. GA October 2026. This is a meta-agent pattern that mirrors what teams are building manually with LangSmith, Langfuse, and similar observability tools, but integrated directly into the Agentforce platform.

Agent Script: Deterministic Control Inside AI Reasoning

One detail that architects should not overlook is Agent Script — Salesforce’s open-source language for agent behavior. Agent Script lets customers combine AI reasoning with deterministic rules for granular control over how agents make decisions and take action.

This addresses one of the most persistent challenges in enterprise agentic AI: the tension between the flexibility of LLM-based reasoning and the predictability that regulated industries require. With Agent Script, an architect can define that certain actions (like issuing a refund over a threshold, or modifying a contract) always follow a deterministic path, while the agent uses AI reasoning for everything else.

The open-source nature of Agent Script is also noteworthy. By making the behavior specification language available outside the Salesforce ecosystem, Salesforce is positioning it as a potential standard for agent guardrails — competing with similar efforts from other vendors and the open-source community.

How This Compares to Competing Enterprise Agent Platforms

The September 2026 announcements don’t exist in a vacuum. Every major cloud provider and enterprise platform is racing to own the enterprise agent layer. Here’s how the Agentforce portfolio stacks up:

Salesforce’s unique advantage is the CRM data layer. Every enterprise agent — regardless of which platform built it — eventually needs access to customer data, sales pipelines, support tickets, and business processes. By positioning the Trusted Enterprise AI Harness as the governance layer for that data, Salesforce is betting that even companies using competing agent platforms will route through Salesforce for the data.

The risk, as ServiceNow’s AI Control Tower and IBM’s approaches demonstrate, is that governance-layer plays are only as strong as the ecosystem they govern. If enterprises standardize on a hyperscaler’s native agent platform (and most will for at least some workloads), they may prefer that platform’s native governance tools over a CRM vendor’s cross-platform harness.

What RPA Teams Should Know

For teams currently running UiPath or Automation Anywhere for procure-to-pay and similar workflows, the Agentforce announcements raise a specific question: when does an agentic CRM agent replace a traditional RPA bot?

The answer, based on the September announcements, is: for CRM-adjacent workflows, increasingly soon. Marshall (the supply chain agent) handles end-to-end back-office processes with deterministic execution and full audit trails — capabilities that overlap directly with what RPA bots do in procurement, order management, and fulfillment. But Marshall operates natively within the Salesforce data model, eliminating the integration layer that RPA bots need to read and write Salesforce data.

For workflows that span systems beyond Salesforce — ERP processing, legacy mainframe interactions, desktop automation — traditional RPA remains the right tool. The convergence point is in the orchestration layer: UiPath’s Autopilot is adding agentic capabilities to RPA, while Agentforce is adding deterministic execution to its agents. Architects should expect these to meet in the middle, and plan integration patterns accordingly.

Dreamforce 2026: What to Watch For

Dreamforce 2026 runs September 15-17 at Moscone Center in San Francisco. Based on pre-event announcements, the key sessions for agentic AI architects are:

• Marc Benioff and Dario Amodei keynote: How Salesforce and Anthropic plan to position Claude, Agentforce, Slack, Data 360, and Headless 360 within a shared enterprise AI architecture
• Agentforce Keynote: “From Fast Start to Real ROI” — the shift from Agentforce pilots to measurable business results
• Agentforce Operations Keynote: Running and scaling agents once they’re in production — this is where the long-horizon runtime’s operational details will likely get deeper coverage
• Trusted Enterprise AI Harness deep-dives: Expect detailed sessions on the AI Control Plane and how it integrates with non-Salesforce agent platforms

Watch for announcements on which additional agents will get the long-horizon runtime (Casey and Fin are likely candidates given the complexity of customer service interactions that span multiple sessions).

Implementation Considerations for Architects

If you’re evaluating Agentforce for your organization, here are the architectural decisions to make now:

Start with the Agent That Matches Your Highest-Volume Workflow

Don’t deploy all seven agents simultaneously. Pick the one that maps to your organization’s highest-volume, most well-defined workflow. Casey (customer service) and Piper (inbound pipeline generation) have the most production evidence and the shortest time-to-value based on early adopter data.

Plan Your Guardrail Architecture Before Deployment

Agent Script gives you deterministic control over critical decision points. Map your compliance requirements, approval thresholds, and escalation rules into Agent Script definitions before the agent goes live — retrofitting guardrails after deployment is significantly harder.

Evaluate the Trusted Enterprise AI Harness Against Your Existing Governance Stack

If you’re already running agents on AWS Bedrock or Azure AI Foundry, the Harness offers a cross-platform governance layer. But it adds a dependency on Salesforce infrastructure for governance — weigh the benefit of unified governance against the coupling this creates.

Understand the Flex Credit Consumption Model

The new editions use Flex Credits for agent actions. Before committing to an edition tier, model your expected AWU volume based on current transaction volumes. The difference between Core (500,000 credits) and Max (2,750,000 credits) is significant, and overages on credit-based models can be expensive.

Plan for Long-Horizon Agent Patterns

Even though only Hunter runs on the long-horizon runtime today, design your agent architecture with persistent state in mind. The pattern will expand, and retrofitting agents for multi-session goals is architecturally harder than building for it from the start.

Frequently Asked Questions

What are Salesforce Agentforce job-ready agents?

Job-ready agents are seven pre-built AI agents — Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin — that ship with domain-specific skills, data models, and integrations already configured for specific business functions like customer service, sales, commerce, HR/IT, supply chain, and pipeline generation. They’re designed to be deployed and customized rather than built from scratch.

How does the Agentforce long-horizon runtime differ from standard agent runtimes?

Standard agent runtimes complete tasks within a single session — when the conversation ends, the agent’s state ends. The long-horizon runtime adds persistent memory, durable execution, and dynamic steering so agents can pursue goals across days and weeks, maintaining context and progress between sessions. Hunter is the first agent to use it, with more planned.

What is the Trusted Enterprise AI Harness and do I need it?

The Trusted Enterprise AI Harness is a six-pillar governance framework (Context, Agency, Action, Governance, Security, Models) with an AI Control Plane that manages agents across multiple vendor platforms. You need it if your organization runs agents from more than one vendor and needs centralized governance, audit trails, and security policies applied consistently across all of them.

How does Claudeforce change the Salesforce-Anthropic relationship?

Claudeforce makes Claude the default reasoning model across multiple Salesforce surfaces (Slack AI, Agentforce Coworker, Headless 360) and adds a Salesforce plugin inside Claude with 37 prebuilt sales skills. It’s the first LLM provider fully integrated within the Salesforce Trust Boundary, meaning Claude can access Customer 360 data under the same security controls as native Salesforce features.

Can Agentforce job-ready agents replace my RPA bots?

For CRM-adjacent workflows (customer service, order management, pipeline generation), the job-ready agents handle many tasks that previously required RPA bots with Salesforce integrations. For workflows spanning non-Salesforce systems — ERP processing, legacy mainframe interactions, desktop automation — traditional RPA remains the better tool. The convergence point is the orchestration layer where both RPA and agentic AI platforms are meeting.

Key Takeaways

• Salesforce shipped seven named, job-ready agents (Casey, Paige, Carter, Hunter, Marshall, Piper, Fin) — six GA now, Hunter in pilot — that compress the time from concept to production deployment
• The long-horizon runtime is the most architecturally significant feature, enabling agents to pursue goals across weeks with persistent memory, durable execution, and dynamic steering
• The Trusted Enterprise AI Harness positions Salesforce as the governance layer for multi-vendor agent deployments, built headlessly with MCP/API access
• Claudeforce makes Anthropic’s Claude the default reasoning model across Agentforce surfaces, with 37 prebuilt sales skills accessible from within Claude’s interface
• New Core/Advanced/Max editions ($195/$395/$550 per seat) bundle Agentforce, Slack, Tableau Next, and security into single SKUs with Flex Credit-based consumption
• Agent Script (open-source) gives architects deterministic control over critical agent decision points — essential for regulated industries
• 7 billion AWUs delivered (3.2B in Q2 alone, 15% CMGR) validates market adoption, though the AWU metric itself remains debated as a value measurement
• For RPA teams: Agentforce agents increasingly overlap with RPA bots for CRM-adjacent workflows; plan integration patterns now rather than after both platforms converge

References

1. Salesforce. “Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work.” September 11, 2026. salesforce.com
2. Salesforce. “Salesforce Introduces the Trusted Enterprise AI Harness.” September 10, 2026. salesforce.com
3. Salesforce. “New Salesforce Editions Bundle Everything Businesses Need for Agentic Transformation.” September 3, 2026. salesforce.com
4. Salesforce & Anthropic. “Salesforce and Anthropic Announce Claudeforce.” August 26, 2026. salesforce.com
5. VentureBeat. “Companies already run 3 agent platforms. Salesforce’s new Enterprise AI Harness wants to govern all of them.” September 2026. venturebeat.com
6. Futurum Group. “Salesforce’s Job-Ready Agents Target Enterprise AI’s Biggest Gap.” September 2026. futurumgroup.com
7. Unite.AI. “Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime.” September 2026. unite.ai
8. Enterprise DNA. “Salesforce Launches 7 Named AI Agents Before Dreamforce.” September 2026. enterprisedna.co
9. Constellation Research. “Salesforce’s Agentic Work Unit: What You Need to Know.” 2026. constellationr.com
10. Salesforce. “The Agentic Work Unit: Converting Raw Intelligence into Real Work.” 2026. salesforce.com