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

Seven Named Agents, One Runtime That Never Forgets

On September 11, 2026 — four days before Dreamforce opens in San Francisco — Salesforce dropped its most consequential Agentforce update yet. Seven named, job-ready AI agents shipped across sales, service, commerce, supply chain, and employee experience. Behind them sits a long-horizon runtime that lets an agent pursue a business goal across days and weeks, not just answer a question in a single chat turn. And underneath all of it, Agent Script, a new open-source declarative language, gives architects deterministic control over what those agents actually do.

The numbers back up the urgency. Salesforce reports it has now delivered 7 billion Agentic Work Units (AWUs) across Agentforce and Slack — with 3.2 billion of those arriving in Q2 alone, a pace that doubled quarter-over-quarter. Early adopter results are just as pointed: Anthropic itself is seeing 79% of customer conversations resolved autonomously by Fin, Salesforce’s customer experience agent. Perk reports 60% of its entire sales pipeline is now built by Hunter, the outbound sales agent still technically in pilot.

For Agentic AI Architects evaluating enterprise agent platforms, this launch changes the calculus. Salesforce is no longer pitching “build your own agent from scratch.” It’s shipping pre-configured agents that know their job on day one, backed by a runtime architecture that rivals what custom-built LangGraph or CrewAI deployments require months of engineering to replicate. Here’s what you need to know.

What Are Salesforce’s Job-Ready Agents?

The core idea is simple but significant: instead of handing enterprises a blank agent canvas and a prompt box, Salesforce now ships agents pre-loaded with the skills, actions, data models, and integration points a specific business role needs. Each agent connects to Customer 360, so it inherits the customer context, business rules, permissions, and security policies already configured in the org.

Enterprises can rename any agent to match their brand (Engine renamed Casey to “Eva,” for example) and tailor its behavior. But the starting point is an agent that already understands the job — not one that needs weeks of prompt engineering to stop hallucinating order statuses.

Here’s the full roster:

Six of the seven are generally available today. Hunter, the outbound sales agent, is in pilot with GA planned for November 2026. Notably, Hunter is also the first agent running on the new long-horizon runtime — the architectural centerpiece of this release.

The Long-Horizon Runtime: Agents That Work for Weeks

This is where the announcement shifts from a product update to an architectural inflection point.

More Read

RPA Center of Excellence: Building and Sustaining Success

RPA and Compliance: Ensuring Regulatory Adherence

IBM watsonx Orchestrate Agentic Control Plane: The Complete Guide for Agentic AI Architects

Most enterprise AI agents today operate in a request-response pattern: a user asks something, the agent reasons over it, returns an answer, and the session ends. Even sophisticated multi-step agents built with frameworks like Pydantic AI or LangGraph typically complete their work within a single session boundary. Context vanishes when the conversation closes.

Salesforce’s long-horizon runtime breaks that pattern. An agent running on this runtime can:

• Accept a high-level business objective (“rescue my at-risk deals before quarter-end”)
• Decompose it into a measurable goal with a multi-step plan
• Execute that plan across days and weeks, incorporating new information as it arrives
• Determine autonomously which tasks it can handle and which require human approval
• Resume, course-correct, or reprioritize as circumstances change

Three capabilities underpin this:

1. Memory

The runtime preserves context and progress across sessions. When a seller closes the laptop on Friday, the agent doesn’t lose what it learned about the deal, the prospect’s last objection, or the competitive intelligence it gathered. Monday morning, it picks up exactly where it left off. This isn’t conversation history in a vector store — it’s structured, persistent state tied to a business objective.

2. Durable Execution

Plans keep running over time. If Hunter is working a deal rescue and a prospect responds to an email at 2 AM, the runtime can process that response, update the plan, and queue the next action — without waiting for a human to re-invoke it. Think of it as a workflow engine with LLM reasoning built in, rather than an LLM with workflow capabilities bolted on.

3. Dynamic Steering

The agent adapts its behavior based on individual user feedback and direction. If a seller says “don’t email this prospect again, call instead,” Hunter adjusts its approach for that specific deal without requiring a global configuration change. The steering is per-user and per-objective, not per-org.

For architects who’ve built long-running agent systems from scratch — stitching together LangGraph state machines, Redis-backed memory, and custom orchestration layers — the significance here is clear. Salesforce is productizing what took custom engineering teams months to build, and embedding it into a platform that already has the CRM data, business rules, and security model in place.

Agent Script: Deterministic Control for Probabilistic Agents

The tension in every enterprise agent deployment is the same: LLMs are probabilistic, but business processes need deterministic guarantees. You can’t have a customer service agent sometimes applying a 20% discount and sometimes applying 50% because the prompt was ambiguous.

Salesforce’s answer is Agent Script, an open-source, declarative language for defining agent behavior. Open-sourced in April 2026 at github.com/salesforce/agentscript, it provides what Salesforce calls the “control plane for agentic decisions.”

Agent Script looks like a hybrid of YAML and Python — indentation-sensitive, with explicit variable types, declared actions, and execution flow logic. Where an LLM’s reasoning is probabilistic, Agent Script is fixed: it declares handoff rules, state management, and execution flow as explicit, readable code. Architects can combine AI reasoning (let the LLM figure out the best response to an angry customer) with deterministic rules (never offer more than 15% off without manager approval, always escalate fraud mentions to a human).

The open-source repo includes the compiler, linter, language server, VS Code extension, and playground. The runtime itself is not yet open-sourced — it runs inside the Agentforce platform — but the language specification is fully open, which matters for portability and community adoption.

For architects who’ve wrestled with fragile prompt engineering as the sole behavior-control mechanism, Agent Script represents a genuinely different approach: code-level determinism wrapping LLM-level intelligence, rather than the other way around.

Multi-Agent Orchestration and the Platform Capabilities

Job-ready agents and the long-horizon runtime are the headline, but several supporting platform capabilities round out the story:

Multi-Agent Orchestration (GA Now)

When a job crosses roles, systems, or stages of the customer journey, Multi-Agent Orchestration routes work across multiple agents so they operate as a coordinated team. Think of a customer return that touches Casey (service), Marshall (supply chain), and Carter (commerce follow-up) — the orchestrator manages the handoffs and shared state. For architects familiar with the A2A (Agent-to-Agent) protocol, this is Salesforce’s proprietary answer to the same problem, though scoped to the Agentforce ecosystem.

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

Employees can teach Agentforce Coworker how to complete a task once, then scale that knowledge across the workforce. This is the “show, don’t program” approach to agent customization — a knowledge worker demonstrates a workflow, and Coworker learns to replicate it. The parallel to UiPath’s Autopilot coding agent is worth noting: both platforms are converging on the idea that the people closest to the work should be able to teach agents directly, without writing code.

Agent Optimizer (GA October 2026)

An agent that optimizes other agents. Agent Optimizer works alongside teams across the agent lifecycle — helping build and refine agents, subagents, and actions, test performance, and analyze session traces to identify improvement opportunities. This meta-agent approach mirrors what observability platforms like Langfuse and LangSmith provide for custom agent stacks, but integrated directly into the Agentforce platform.

Claudeforce: When the #1 AI Meets the #1 AI CRM

Two weeks before the job-ready agents launch, Salesforce and Anthropic announced Claudeforce on August 26, 2026 — the first time Salesforce has ever added its “force” suffix to another company’s product. The partnership runs in two directions:

• Claude into Salesforce: Claude becomes a reasoning model option within the Agentforce platform, bringing Anthropic’s extended thinking and tool-use capabilities to enterprise agent workflows.
• Salesforce into Claude: A plugin with 37 prebuilt sales skills enables sellers and agents to reason over live revenue context, automate pipeline updates, and take governed action directly from Claude’s interface.

Claudeforce is currently available to a select group of pilot customers, with a preview launch planned for late 2026. For architects, the subtext is strategic: Salesforce is positioning Agentforce as model-agnostic. Your agents can reason with Claude, with GPT-6 Astra, or with Salesforce’s own models — the platform’s value is in the data, orchestration, and governance layers, not in any single LLM.

Pricing: What Agentforce Actually Costs

Salesforce runs three pricing models for Agentforce in 2026:

The AWU metric (Agentic Work Unit) is a disclosure and reporting metric, not a pricing unit. Salesforce defines one AWU as one discrete task accomplished by an AI agent — a prompt processed, a reasoning chain completed, or a tool invoked. It appears in earnings reports and marketing materials, but there’s no published conversion between AWUs and Flex Credits.

For architects doing cost modeling: a typical customer service conversation might consume 3-8 agent actions (40-160 Flex Credits, or $0.20-$0.80), making the $2/conversation flat rate simpler but potentially more expensive for short interactions. The Flex Credit model rewards optimization — fewer, more targeted agent actions per resolution mean lower cost per outcome.

Early Customer Results: What the Numbers Actually Show

Salesforce published six customer case studies alongside the launch. Here’s what they reported:

A few observations for architects evaluating these numbers. First, the resolution rates (50-90%) are impressive but vary significantly by use case complexity. Service desk inquiries (Autism Queensland’s 70%) are typically more structured than sales conversations (Perk’s 60% pipeline generation). Second, the Hibbett deployment timeline — six weeks to 90% journey coverage — suggests that the “job-ready” framing isn’t just marketing. Pre-configured agents with the right data model genuinely compress time-to-value compared to building from scratch. Third, the Anthropic result is meta: an AI company using Salesforce’s AI agent to handle its own customer conversations, and seeing 79% autonomous resolution.

What This Means for Agentic AI Architects

This launch crystallizes a strategic split that architects need to understand:

Build vs. Buy Is Now a Three-Way Decision

Until this release, the enterprise agent landscape offered two paths: build custom agents with open-source frameworks (LangGraph, CrewAI, AutoGen) or configure agents on a platform (Agentforce, Microsoft Agent Framework, IBM watsonx Orchestrate). Salesforce’s job-ready agents introduce a third option: adopt pre-built agents that already know the job and customize from there.

For organizations deeply embedded in the Salesforce ecosystem, the adoption path is obvious. For those running multi-vendor stacks, the question becomes whether Salesforce’s agent governance (Agent Script, permissions, audit trails) justifies platform lock-in versus the flexibility of open-source orchestration with a tool like Orca ADE.

Long-Horizon Runtime Sets a New Bar

The ability for an agent to pursue a goal across days and weeks — with persistent memory, durable execution, and dynamic steering — isn’t something most custom agent deployments have today. Architects building with LangGraph or similar frameworks can replicate this, but it requires significant infrastructure work: persistent state stores, workflow engines, checkpoint/resume logic, and user-specific behavior adaptation. Salesforce is offering it as a platform capability, which raises the minimum viable feature set for any competing enterprise agent platform.

Expect Microsoft’s Agent Framework and Google’s Gemini Enterprise Agent Platform to respond with similar long-horizon capabilities within the next two quarters. The pattern of agents working across extended timeframes, not just single conversations, is becoming table stakes for enterprise deployments.

Agent Governance Is the Real Differentiator

As enterprises move from pilot to production, the conversation is shifting from “can the agent do the task?” to “can we prove the agent did the task correctly, within policy, with an audit trail?” Salesforce’s combination of Agent Script (deterministic behavior control), Customer 360 permissions (inherited security), and Marshall’s audit-trail-first design for supply chain operations reflects this shift. Architects evaluating platforms should weight governance capabilities — audit logs, deterministic guardrails, human-in-the-loop controls, compliance reporting — at least as heavily as raw agent intelligence. This parallels the enterprise agent governance frameworks we’ve seen from SAP AI Agent Hub and ServiceNow AI Control Tower.

Dreamforce 2026: What to Watch

Dreamforce 2026 runs September 15-17 at the Moscone Center in San Francisco under the theme of becoming an “Agentic Enterprise.” With 1,600+ breakout sessions, 50+ keynotes, 150+ hands-on trainings, and 240+ community roundtables, the event will serve as the live showcase for everything announced here.

Key sessions to watch:

• Marc Benioff’s opening keynote — expect the “Agentic Enterprise” vision statement, likely with live demos of Hunter’s long-horizon runtime and at least one Claudeforce integration showcase.
• “Trust Across Agents, Data, and Platforms” security keynote (Tuesday, September 15, 2:15 PM PT) — with 30+ security-focused sessions, agent governance will be a major theme.
• Agent Script deep dives — look for technical sessions on how to combine deterministic Agent Script rules with LLM reasoning in production deployments.

A free virtual program on Salesforce+ runs September 15-18 (one day beyond the in-person event) with 400+ sessions.

Frequently Asked Questions

What are Salesforce Agentforce job-ready agents?

Job-ready agents are pre-configured AI agents that ship with the skills, actions, and data models required for a specific business role — customer service (Casey), IT/HR (Paige), commerce (Carter), outbound sales (Hunter), supply chain (Marshall), inbound pipeline (Piper), and customer experience (Fin). They connect to Customer 360 and inherit existing business rules and security, so enterprises can deploy them without building from scratch.

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

Standard AI agents operate within a single conversation session — when the chat ends, context is lost. The long-horizon runtime enables agents to pursue business objectives across days and weeks, with persistent memory, durable execution that continues between sessions, and dynamic steering that adapts to individual user feedback. Hunter is the first agent using this runtime.

What is Salesforce Agent Script?

Agent Script is an open-source, declarative language for defining deterministic agent behavior. It lets architects combine LLM reasoning with fixed business rules (handoff conditions, approval thresholds, escalation triggers) in readable code rather than fragile prompt engineering. The compiler, linter, and VS Code extension are open-source; the runtime runs inside Agentforce.

How much does Salesforce Agentforce cost in 2026?

Salesforce offers three pricing models: $2 per conversation (flat rate), Flex Credits at $500 per 100,000 credits (with actions consuming 20-60 credits each), and per-user licenses starting at $125/month for Agentforce Coworker. The AWU metric is used for reporting, not billing.

What is Claudeforce?

Claudeforce is the strategic partnership between Salesforce and Anthropic, announced August 26, 2026. It brings Claude as a reasoning model into Agentforce and embeds Salesforce into Claude via a plugin with 37 prebuilt sales skills. It’s currently in pilot with preview launch planned for late 2026.

Key Takeaways

• Seven job-ready agents ship pre-configured for sales, service, commerce, supply chain, IT/HR, and CX — six are GA now, Hunter (outbound sales) is in pilot with November 2026 GA.
• The long-horizon runtime is the architectural headline: agents can pursue business goals across days and weeks with persistent memory, durable execution, and dynamic steering.
• Agent Script provides deterministic, code-level control over agent behavior — an open-source answer to the fragile-prompt-engineering problem.
• 7 billion AWUs delivered across Agentforce and Slack (3.2B in Q2 alone) signals production-scale adoption, not pilot-phase experimentation.
• Claudeforce makes Agentforce model-agnostic — Claude, GPT-6 Astra, and Salesforce’s own models can all serve as reasoning engines.
• Pricing runs $2/conversation, $0.005/Flex Credit, or $125+/month per user — architects should model cost based on actions-per-resolution.
• Enterprise agent governance (Agent Script + permissions + audit trails) is becoming the real platform differentiator, not raw LLM intelligence.
• Dreamforce 2026 (September 15-17) will showcase these capabilities live — watch for long-horizon runtime demos and Claudeforce integration previews.

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 and Anthropic Announce Claudeforce: The #1 AI Meets the #1 AI CRM.” August 26, 2026. salesforce.com
3. Salesforce. “Agent Script: The Control Plane for Agentic Decisions.” 2026. salesforce.com
4. Salesforce. “Agentforce Script — Open-Source Repository.” GitHub. github.com/salesforce/agentscript
5. Unite.AI. “Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime.” September 11, 2026. unite.ai
6. Enterprise DNA. “Salesforce Launches 7 Named AI Agents Before Dreamforce.” September 2026. enterprisedna.co
7. ppc.land. “Salesforce agents gain a runtime that pursues goals over weeks, not chats.” 2026. ppc.land
8. CNBC. “Salesforce, Anthropic expand partnership as Benioff responds to ‘SaaSpocalypse’ concerns.” August 26, 2026. cnbc.com
9. Salesforce. “Agentforce Pricing — Flex Credits.” 2026. salesforce.com/agentforce
10. MarketScale. “Dreamforce 2026 puts the agentic enterprise on trial in San Francisco this September.” 2026. marketscale.com