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

Seven billion. That is the number of Agentic Work Units Salesforce says it has delivered across Agentforce and Slack — including 3.2 billion in Q2 2026 alone. On September 11, 2026, the company made its clearest move yet to translate that operational volume into a repeatable enterprise pattern: a portfolio of seven named, job-ready AI agents, a long-horizon runtime that lets agents pursue goals across days and weeks instead of single conversations, and a set of platform capabilities — multi-agent orchestration, AI Skills, and Agent Optimizer — that together represent the most significant Agentforce expansion since the platform’s Dreamforce 2024 launch.

For agentic AI architects evaluating where enterprise agent platforms are headed, this announcement answers a question the market has been circling for months: what happens after you prove AI can answer questions? Salesforce’s answer is that AI must do the work — across sales, service, commerce, employee experience, and back-office operations — and that it needs to do so with the kind of persistence, coordination, and continuous improvement that single-turn chatbot architectures simply cannot deliver.

This guide breaks down every component of the September 11 announcement, examines the architecture and pricing, compares Agentforce to competing enterprise agent platforms, and maps what this means for the architects and developers who will actually deploy these systems.

What Are Agentforce Job-Ready Agents?

The core idea behind job-ready agents is deceptively simple: instead of requiring every enterprise to build agents from scratch — defining skills, actions, data models, and guardrails one by one — Salesforce now ships agents pre-configured for the jobs where AI can deliver immediate business impact. Each agent comes with domain-specific skills, pre-built actions, and data models tailored to its function, and each can be customized to match how a specific company operates, including giving it a brand-aligned name.

This is a meaningful architectural shift. Prior to this release, Agentforce customers worked with a general-purpose agent builder. The new portfolio introduces a role-based agent taxonomy — agents named and designed around business functions rather than technical capabilities. The distinction matters because it moves the starting point from “what can this agent do?” to “what job does this agent fill?”

Every job-ready agent connects to Customer 360, Salesforce’s unified data layer. This means agents inherit the customer context, business processes, permissions, and security rules already configured in Salesforce. They do not operate in isolation — they operate within the enterprise’s existing CRM fabric.

The Seven Named Agents: Capabilities, Domains, and GA Status

Salesforce introduced seven named agents, six of which are generally available today. The following table summarizes each agent’s role, domain, and current status:

Casey: The Multi-Channel Help Agent

Casey handles customer service across voice, SMS, WhatsApp, and web chat — the four channels where most B2C service volume concentrates. Pre-built skills cover the tasks that account for the majority of Tier 1 support tickets: FAQ resolution, product returns, account management, and escalation to human agents when the issue exceeds Casey’s confidence threshold. The multi-channel design is significant because it means a single agent configuration covers the entire service surface rather than requiring separate bot deployments per channel.

More Read

AI Agent Control Planes: The 2026 Enterprise Governance Guide

OpenAI DevDay 2026: Dots, GPT-6.1 Sol, and the Agents API — The Complete Guide for Agentic AI Architects

Oracle AI Agent Studio for Industry Applications: The Complete Guide for Agentic AI Architects (2026)

Hunter: The Long-Horizon Sales Agent

Hunter is the most architecturally interesting agent in the portfolio because it is the first to run on Salesforce’s new long-horizon runtime (covered in detail below). Unlike Casey or Carter, which resolve discrete interactions, Hunter operates on a sales timeline — weeks to months of research, outreach, follow-up, and deal management. A seller can assign Hunter a strategic goal (“rescue at-risk deals before quarter-end”), and Hunter decomposes that into a plan, executes autonomously where authorized, and surfaces decisions that require human approval. Hunter is currently in pilot, with general availability planned for November 2026.

Fin: The CX Agent with Custom Models

Fin stands out because of its model architecture. Rather than relying solely on general-purpose LLMs, Fin is powered by two specialized components: Operator, a customer operations agent that handles workflow orchestration, and Fin Apex, a set of custom models specifically trained for customer experience use cases. This dual-layer architecture — a general orchestrator backed by domain-specific models — is a pattern that is becoming increasingly common in production agent systems, and it explains why Salesforce reports that 79% of Anthropic’s conversations handled by Fin are resolved autonomously.

Marshall: Deterministic Execution for the Back Office

Marshall is Salesforce’s answer to a persistent enterprise concern: what about processes where AI creativity is a liability? Supply chain operations, procurement workflows, and back-office processes often require deterministic execution — every step auditable, every action traceable, no probabilistic deviation. Marshall combines agent orchestration with deterministic execution paths and produces a full audit record of every action, making it viable for regulated industries and compliance-sensitive workflows.

The Long-Horizon Runtime: Agents That Work for Weeks, Not Minutes

The most significant technical announcement in the September 11 release is the long-horizon runtime — a new execution layer that enables Agentforce agents to pursue goals across days and weeks rather than completing only a single task or conversation.

This addresses what may be the single biggest limitation of current enterprise agent deployments: temporal scope. Most production agents today are reactive and ephemeral — they respond to a trigger, execute a sequence, and terminate. They do not remember what they did yesterday, they cannot adapt a plan mid-execution based on new information, and they certainly cannot coordinate a multi-week campaign with a human collaborator. The long-horizon runtime changes each of these constraints.

Three Pillars of the Long-Horizon Runtime

Salesforce describes the runtime as resting on three capabilities:

1. Memory. The runtime preserves context and progress across sessions. When Hunter researches an account on Monday, identifies a risk signal on Wednesday, and drafts an outreach on Friday, each session has access to what came before. This is not conversational memory (retaining chat history within a session) — it is persistent operational memory that survives across independent execution cycles. For architects familiar with agentic frameworks like LangGraph or CrewAI, think of this as durable state management at the platform level, without requiring custom persistence layers.

2. Durable execution. Plans keep running over time. If an agent’s execution is interrupted — by a system restart, a timeout, or a dependency on an external event (waiting for a prospect to reply to an email) — the runtime can resume from the last checkpoint. This is conceptually similar to workflow durability in Temporal or AWS Step Functions, but applied to an agent’s goal-pursuit loop rather than a deterministic workflow.

3. Dynamic steering. The agent adapts its behavior based on individual user feedback and direction. If a seller tells Hunter “don’t email this prospect before Thursday” or “focus on the enterprise tier first,” Hunter incorporates that instruction into its ongoing plan without requiring a full re-planning cycle. This is a practical acknowledgment that real sales work involves continuous micro-adjustments that cannot all be anticipated at plan-creation time.

Architectural Implications for the Agentic AI Community

The long-horizon runtime positions Salesforce closer to what the open-source agentic community has been building with frameworks like Orca ADE and LangGraph’s persistence layers — but with a critical difference: it ships as a managed, platform-integrated capability rather than a framework you assemble yourself. For enterprise buyers, this is the value proposition. For architects evaluating build-vs-buy, it raises the question of whether the flexibility of open-source agent orchestration is worth the infrastructure overhead when the CRM data already lives in Salesforce.

Platform Capabilities: Multi-Agent Orchestration, AI Skills, and Agent Optimizer

Beyond the named agents, Salesforce announced three platform capabilities that determine how agents are taught, coordinated, and improved.

Multi-Agent Orchestration (GA Now)

Multi-Agent Orchestration routes work across agents so they can operate as one coordinated team when a job crosses roles, systems, or stages of the customer journey. In practice, this means an orchestrator agent connects to multiple specialized agents within an org and presents a single unified interface to the user. A customer service interaction that starts with Casey (help agent) but requires input from Marshall (supply chain) and Carter (commerce) can be routed seamlessly without the customer knowing multiple agents are involved.

This is Salesforce’s production implementation of the multi-agent coordination patterns that the broader agentic AI community has been developing through protocols like A2A (Agent-to-Agent) and MCP (Model Context Protocol). The key difference is that Salesforce’s orchestration is tightly coupled to Customer 360 data and Salesforce’s permission model, which simplifies governance but limits portability.

AI Skills in Agentforce Coworker (Pilot Now, GA October 2026)

AI Skills introduces a teach-once-scale-everywhere pattern for Agentforce Coworker, Salesforce’s employee-facing agent. An employee teaches Coworker how to complete a task — say, preparing a quarterly business review deck from CRM data — and that skill is then available to every Coworker instance across the workforce and across interfaces (Slack, Salesforce UI, mobile). This is Salesforce’s version of the reusable agent skill pattern, and it directly addresses the biggest bottleneck in enterprise agent adoption: the cost of encoding organizational knowledge into agent behavior.

Agent Optimizer (GA October 2026)

Agent Optimizer is an agent that optimizes other agents. It works alongside teams across the agent lifecycle — helping build and refine agents, subagents, and actions, test performance, and analyze session traces to identify what to improve. This closes the observability gap that plagues most enterprise agent deployments: once agents are live, how do you systematically identify where they fail, why they fail, and what to change? Agent Optimizer treats agent improvement as a continuous, data-driven process rather than a periodic manual review.

Agent Script: Salesforce’s Open-Source Agent Orchestration Language

One of the more consequential but less-discussed aspects of Salesforce’s agentic strategy is Agent Script, an open-source language for defining agent behavior that Salesforce released under the Apache 2.0 license. The full toolchain — parser, linter, compiler, Language Server Protocol (LSP) support, and editor integrations — is available on GitHub.

Agent Script lets customers combine AI reasoning with deterministic rules for granular control over how agents make decisions and take action. In practice, this means you can write explicit rules (“never offer a discount above 15% without manager approval”) alongside AI-driven decision steps (“determine which of these three responses best matches the customer’s intent”), and the agent follows both.

For architects coming from RPA backgrounds, Agent Script occupies the space between process definition (like UiPath’s XAML workflows) and prompt engineering. It is more structured than a system prompt and more flexible than a deterministic workflow — which is exactly the sweet spot enterprise agentic deployments need. The open-source release is strategically important because it creates a portable standard: even if a customer moves away from Salesforce, their agent behavior definitions are not locked in.

Claudeforce: The Anthropic Partnership

Two weeks before the job-ready agents announcement, on August 26, 2026, Salesforce and Anthropic announced Claudeforce — a deep integration that brings Claude’s reasoning capabilities into the Agentforce platform and brings Salesforce’s data and actions into Claude.

Claudeforce launches with Salesforce in Claude, a plugin with 37 prebuilt sales skills. These skills — including meeting prep, deal health review, and pipeline review — enable sellers and agents to reason over live CRM data, automate pipeline updates, and take governed action directly from the Claude interface. The integration is bidirectional: Claude moves into Salesforce as a reasoning model powering Agentforce agents, and Salesforce moves into Claude as a structured data and action layer.

Select pilot customers have access now, with open beta expected in September 2026 and additional prebuilt skills launching from late 2026. For architects, the Claudeforce partnership signals that Salesforce views model diversity — not single-vendor lock-in — as the path forward for enterprise reasoning. Agentforce already supports multiple LLM providers; Claudeforce makes the most prominent alternative model a first-class citizen in the platform.

Customer Results: What the Numbers Actually Say

Salesforce cited early results from named customers running these agents. The numbers are worth examining critically — both for what they demonstrate and for the context they leave out.

Several observations for practitioners:

The resolution rates vary significantly by domain. Casey-based deployments (customer service) show 50% resolution, while Fin-based deployments (customer experience with custom models) reach 79%. The difference likely reflects Fin’s specialized model architecture (Fin Apex) versus Casey’s general-purpose approach — a data point supporting the argument that domain-specific fine-tuning materially improves agent performance.

The Hibbett deployment timeline is notable. Going live in six weeks and handling 90% of core shopper journeys suggests that commerce is the fastest path to production for these agents — likely because shopper interactions are more structured and predictable than sales or service conversations.

The Piper metric is volume, not conversion. A 4x increase in conversation volume is a leading indicator, but what matters for pipeline generation is conversion rate. Asana’s average deployment time of 45 days is useful context for planning, but architects should push for conversion data before building business cases around inbound pipe gen agents.

Agentforce Pricing in 2026: What It Actually Costs

Agentforce pricing has evolved into two primary models, and the distinction matters for deployment planning.

Conversation-Based Pricing

Under the original model introduced at Dreamforce 2024, customer-facing agents are priced at $2 per conversation. A conversation is a discrete customer interaction — a service ticket resolved, a shopper guided to checkout, a lead qualified. This model is straightforward for budgeting but can become expensive at high volumes.

Flex Credits

The newer Flex Credits model prices agent actions at a more granular level. A pack of 100,000 Flex Credits costs $500, and each standard action (updating a record, answering from a knowledge base, summarizing a case) consumes 20 credits — working out to approximately $0.10 per action. Complex actions consume more credits. This model gives architects finer control over cost optimization by designing agents that minimize unnecessary actions.

Agentic Work Units: A Metric, Not a Price

An important clarification: the Agentic Work Unit (AWU) that Salesforce uses in earnings reporting (7 billion delivered, 3.2 billion in Q2) is a platform-level disclosure metric, not a pricing unit. One AWU represents one discrete task accomplished by an AI agent. Salesforce has published no price per AWU and no conversion between AWUs and Flex Credits. If you see a blog quoting a dollar figure per AWU, that is someone’s arithmetic, not Salesforce’s price.

Total Cost Considerations

The per-conversation and per-action pricing covers agent execution only. Total cost of ownership includes the underlying Salesforce platform licenses (Sales Cloud, Service Cloud, Commerce Cloud, etc.), Data Cloud for the unified data layer, any custom model training (relevant for Fin Apex-style deployments), and the implementation partner costs for initial configuration and ongoing optimization. For enterprise deployments, the agent execution cost is often a fraction of the total investment — a pattern familiar to anyone who has deployed RPA at scale, where bot license costs are dwarfed by development, maintenance, and infrastructure spend.

How Agentforce Compares: Copilot Studio, ServiceNow, and UiPath

Salesforce’s job-ready agents enter a market where Microsoft, ServiceNow, and UiPath all have competing enterprise agent platforms. Here is how they compare across the dimensions that matter most to architects:

When to choose Agentforce: Your organization’s customer data, sales workflows, and service operations live in Salesforce. You want pre-built agents that are already wired into your CRM data model. You need long-horizon agents for sales or pipeline management. The Claudeforce partnership matters to your AI strategy.

When to choose Copilot Studio: Your organization runs on Microsoft 365 and Dynamics 365. Employee productivity agents (Copilot in Teams, Outlook, Word) are a higher priority than CRM agents. Your Azure infrastructure is already established. Read more in our guide to Microsoft Agent Framework 1.0 and Foundry Agent Service.

When to choose ServiceNow: ITSM is your core use case. You need cross-platform agent governance as a primary capability. Your enterprise already runs significant ServiceNow deployments. See our deep dive on ServiceNow AI Control Tower for enterprise agent governance.

When to choose UiPath: Your automation strategy is RPA-first and you are extending into agentic workflows. You need UI automation (screen scraping, desktop interaction) alongside AI agent capabilities. Your compliance requirements demand deterministic execution with full audit trails. See our coverage of UiPath Autopilot as a coding agent.

What This Means for Agentic AI Architects

The September 11 announcement crystallizes several trends that architects should plan around:

1. The Shift from Agent Builders to Agent Libraries

Salesforce is not the only vendor moving from “here’s a framework to build agents” to “here are agents built for your job.” Microsoft’s Copilot agents, ServiceNow’s domain-specific agents, and IBM’s watsonx Orchestrate are all converging on this pattern. For architects, this means the competitive advantage shifts from building agents to selecting, customizing, and orchestrating agents — a skill set closer to enterprise architecture than software engineering.

2. Long-Horizon Execution Is the Next Frontier

Most production agents today are stateless and single-turn. The long-horizon runtime’s memory, durable execution, and dynamic steering capabilities represent where the entire industry is heading. Open-source frameworks like LangGraph and Orca ADE are building similar persistence layers, but Salesforce’s managed offering lowers the infrastructure bar for enterprises that do not want to operate their own agent state management. Architects should start designing for multi-session agent lifecycles now — even if the initial deployment is single-turn.

3. Multi-Agent Orchestration Requires a Governance Model

When one agent handles a customer interaction, governance is straightforward: define the agent’s permissions, guardrails, and escalation paths. When seven agents coordinate across a customer journey, governance becomes a systems problem. Who is responsible when Marshall’s supply chain decision conflicts with Hunter’s sales commitment? How are cross-agent permissions managed? What does the audit trail look like when three agents contributed to a single outcome? These are the questions architects need to answer before deploying multi-agent orchestration, and they mirror the governance challenges that A2A and MCP protocols are trying to solve at the industry level.

4. Open-Source Agent Languages Will Proliferate

Agent Script’s Apache 2.0 release is a strategic play to establish a portable standard for agent behavior definition. If Agent Script gains adoption beyond the Salesforce ecosystem — which is the purpose of open-sourcing it — it could become to agent orchestration what BPMN became to business process management: an imperfect but widely understood common language. Architects should evaluate Agent Script alongside other emerging agent definition standards (OpenAI’s function-calling schema, Anthropic’s tool-use protocol, the MCP specification) and start forming a view on which will survive the standards shakeout.

Frequently Asked Questions

What are Salesforce Agentforce job-ready agents?

Job-ready agents are pre-built AI agents designed for specific business functions — sales, service, commerce, employee experience, and back-office operations. Each ships with domain-specific skills, actions, and data models, and connects to Salesforce Customer 360 for access to existing business data. Seven named agents were announced on September 11, 2026: Casey (help), Paige (IT/HR), Carter (commerce), Hunter (sales), Marshall (supply chain), Piper (inbound pipeline), and Fin (customer experience).

How does the Agentforce long-horizon runtime work?

The long-horizon runtime enables agents to pursue goals across days and weeks rather than completing a single task or conversation. It relies on three capabilities: memory (preserving context across sessions), durable execution (keeping plans running over time with checkpoint-based resumption), and dynamic steering (adapting behavior based on user feedback). Hunter, the outbound sales agent, is the first agent running on this runtime.

How much does Agentforce cost in 2026?

Agentforce offers two primary pricing models: $2 per conversation for customer-facing agents, and Flex Credits at $500 per 100,000 credits (approximately $0.10 per standard action). Agentic Work Units (AWUs) are a reporting metric, not a pricing unit — Salesforce has published no price per AWU. Total cost includes underlying Salesforce platform licenses, Data Cloud, and implementation costs.

How does Agentforce compare to Microsoft Copilot Studio?

Agentforce is CRM-native and strongest for sales, service, and commerce use cases where data lives in Salesforce. Copilot Studio is integrated with Microsoft 365 and Azure, making it the natural choice for organizations whose daily workflows center on Teams, Outlook, and Dynamics 365. The right choice typically follows where an organization’s data and workflows already live, not the AI feature list. Agentforce’s long-horizon runtime and Agent Script open-source language are differentiators without direct Copilot Studio equivalents.

What is Agent Script and why does it matter?

Agent Script is an open-source language (Apache 2.0) for defining agent behavior. It lets enterprises combine AI reasoning with deterministic rules — specifying exactly how agents make decisions and take actions. The full toolchain (parser, linter, compiler, LSP, editor integrations) is on GitHub. It matters because it creates a portable agent behavior standard that is not locked to the Salesforce platform, giving enterprises confidence that their agent definitions remain usable if they change platforms.

Key Takeaways

• Seven named, job-ready agents cover sales (Hunter, Piper), service (Casey, Fin), commerce (Carter), employee experience (Paige), and back-office operations (Marshall) — six are GA now, Hunter enters GA in November 2026.
• The long-horizon runtime is the most architecturally significant feature — agents that pursue goals over weeks with persistent memory, durable execution, and dynamic steering represent a fundamental shift from single-turn agent patterns.
• Multi-Agent Orchestration is GA now and enables coordinated agent teams, but requires architects to solve cross-agent governance before deployment.
• Agent Script (Apache 2.0) creates a portable standard for agent behavior definition, bridging AI reasoning and deterministic rules.
• Claudeforce deepens the Anthropic partnership with 37 prebuilt sales skills and bidirectional CRM-to-Claude integration.
• Pricing spans two models: $2/conversation (fixed) or Flex Credits at ~$0.10/action (granular). AWUs are a reporting metric, not a pricing unit.
• Platform choice follows data: choose Agentforce when CRM data lives in Salesforce, Copilot Studio for M365/Azure shops, ServiceNow for ITSM-first, UiPath for RPA-extending-into-agentic.
• 7 billion AWUs delivered (3.2B in Q2 2026) demonstrates production-scale adoption, but architects should push for outcome metrics (conversion rates, CSAT) beyond resolution percentages.

References

1. Salesforce. “Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work.” September 11, 2026.
2. Unite.AI. “Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime.” September 11, 2026.
3. Futurum Group. “Salesforce’s Job-Ready Agents Target Enterprise AI’s Biggest Gap.” September 2026.
4. Enterprise DNA. “Salesforce Launches 7 Named AI Agents Before Dreamforce.” September 2026.
5. Technobezz. “Salesforce Adds Seven AI Agents to Agentforce.” September 2026.
6. Salesforce. “Salesforce and Anthropic Announce Claudeforce.” August 26, 2026.
7. CNBC. “Salesforce, Anthropic expand partnership as Benioff responds to ‘SaaSpocalypse’ concerns.” August 26, 2026.
8. Salesforce. “What are Agentic Work Units (AWU)? Measuring AI Value.” 2026.
9. Salesforce. “Agent Script — GitHub Repository.” Apache 2.0 License.
10. Smartbridge. “Salesforce Agentforce vs Microsoft Copilot Studio 2026 Comparison.” 2026.

Next steps: Run this article through the rpabotsworld-seo skill for a full SEO audit before publishing, and use rpabotsworld-visuals for a matching architecture diagram and YouTube thumbnail.