Every Company Wants AI That Does the Work — Salesforce Just Shipped Seven Agents That Try
On September 11, 2026, Salesforce did something no other enterprise platform vendor has done at this scale: it shipped seven named, job-ready AI agents — each pre-built for a specific business function, each generally available (or close to it), and each backed by a new long-horizon runtime that lets agents pursue goals across days and weeks instead of completing a single chat turn and forgetting everything.
- Every Company Wants AI That Does the Work — Salesforce Just Shipped Seven Agents That Try
- The Seven Named Agents: What Each One Does
- Long-Horizon Runtime: Agents That Remember Across Days and Weeks
- AI Skills in Agentforce Coworker: Scaling Institutional Knowledge
- Multi-Agent Orchestration: Agents Working as a Coordinated Team
- Agent Script: Open-Source Deterministic Control
- Agent Optimizer: An Agent That Improves Other Agents
- Claudeforce: Anthropic Meets Customer 360
- The Numbers: 7 Billion Agentic Work Units and What They Mean
- Where Salesforce Sits in the Agentic Platform Landscape
- What Agentic AI Architects Should Do Next
- Frequently Asked Questions
- What are Salesforce Agentforce job-ready agents?
- How does the Agentforce long-horizon runtime differ from standard AI agents?
- What is Agent Script and is it truly open source?
- How much does Salesforce Agentforce cost?
- Can Agentforce agents work with non-Salesforce systems?
- Key Takeaways
- References
The announcement also introduced AI Skills in Agentforce Coworker, a generally available Multi-Agent Orchestration layer, and Agent Optimizer — an agent that improves other agents. Combined with the Claudeforce partnership with Anthropic announced two weeks earlier, Salesforce is making its clearest bet yet: the future of enterprise software is not dashboards and workflows — it is agents that do the work humans used to do, grounded in Customer 360 data.
For Agentic AI Architects evaluating where Salesforce now sits in the vendor landscape — and whether its approach to agent persistence, orchestration, and deterministic control actually holds up — here is everything you need to know.
The Seven Named Agents: What Each One Does
Instead of shipping a generic “build your own agent” toolkit and leaving enterprises to figure out the rest, Salesforce took a product-management approach: identify the seven highest-impact job categories where AI can deliver measurable results today, then ship purpose-built agents for each. Every agent connects to Customer 360, respects existing business rules and permissions, and can be renamed and tailored to match a company’s brand.
| Agent Name | Function | What It Does | Status |
|---|---|---|---|
| Casey | Customer Service | Resolves customer issues across voice, SMS, WhatsApp, and web chat. Pre-built for FAQs, returns, account management, and human escalation. | GA now |
| Paige | IT & HR Service | Resolves internal employee requests across Slack, portals, and existing tools. | GA now |
| Carter | Commerce | Helps shoppers discover products, compare options, answer questions, and convert with in-chat checkout. | GA now |
| Hunter | Outbound Sales | Works a sales pipeline from research to outreach, collaborating with sellers over weeks and months. First agent on the long-horizon runtime. | Pilot now; GA Nov ’26 |
| Marshall | Supply Chain | Orchestrates end-to-end back-office processes with deterministic execution and full audit trail. | GA now |
| Piper | Inbound Pipe Gen | Engages, qualifies, and converts inbound leads across websites and inboxes for B2B teams. Powered by Qualified. | GA now |
| Fin | Customer Experience | Resolves complex CX workflows across every channel. Powered by Operator and Fin Apex custom models. | GA now |
What makes this roster notable is not just the breadth — spanning service, sales, commerce, HR/IT, and supply chain — but the specificity. Each agent ships with pre-built skills, actions, and data models for its job. Casey does not need to be taught what a return flow looks like. Marshall does not need custom actions to generate audit records. The premise is that enterprises should be customizing agents for their business, not building the foundational capabilities from scratch.
Early Customer Results
Salesforce reported measurable outcomes from early deployments across the agent portfolio:
- Engine: 50% of chat inquiries fully resolved by its help agent, Eva
- Perk: 60% of sales pipeline built by its outbound sales agent (Hunter)
- Autism Queensland: 70% of administrative requests resolved by Paige
- Hibbett: 90% of core shopper journeys handled by Hibbett AI, live in six weeks
- Asana: 4x conversation volume driven by Piper, deployed in 45 days on average
- Anthropic: 79% of conversations resolved autonomously by Fin
These numbers deserve scrutiny — vendor-reported metrics always do — but the consistency across different business functions and company sizes suggests the job-ready approach is landing faster than build-from-scratch alternatives.
Long-Horizon Runtime: Agents That Remember Across Days and Weeks
This is the most architecturally significant piece of the announcement, and the one most likely to reshape how Agentic AI Architects think about agent design patterns.
The core problem: most enterprise AI agents today operate in a request-response loop. A user asks a question, the agent reasons over it, returns an answer, and the session ends. Even agents with tool-calling capabilities typically complete their work within a single interaction. But the highest-value enterprise work — closing a deal, resolving a complex compliance issue, managing a procurement cycle — unfolds over days, weeks, or months.
Salesforce’s long-horizon runtime addresses this with three capabilities working together:
1. Memory
The runtime preserves context and progress across sessions. When a seller asks Hunter to rescue at-risk deals before quarter-end, Hunter does not start from zero each time the seller returns. It remembers what it learned about each deal, what outreach it sent, what responses came back, and what adjustments it made to its plan. This is not simple chat history retrieval — it is structured memory that the agent uses to inform ongoing decision-making.
2. Durable Execution
Plans keep running even when no one is interacting with the agent. If Hunter determines that a follow-up email should go out on Thursday and a pricing proposal needs review on Friday, those tasks persist in the execution runtime. The agent can resume, retry on failure, or course-correct when circumstances change — without requiring a human to re-initiate the workflow.
For architects familiar with workflow orchestration patterns, this is conceptually similar to what durable execution engines like Temporal provide for microservices — but applied to agent behavior rather than service choreography.
3. Dynamic Steering
The agent adapts its behavior based on individual user feedback and direction during the course of a long-running plan. If a seller tells Hunter “don’t contact the VP of Engineering — go through procurement instead,” Hunter incorporates that constraint into its ongoing plan without requiring a full re-prompt or plan reset.
Together, these three capabilities represent a meaningful shift from the conversational agent paradigm toward what Salesforce is calling goal-directed agency. The agent does not just respond to queries — it pursues objectives, maintains state, and adjusts strategy over time.
Architect’s note: Hunter is currently the only agent running on the long-horizon runtime. Salesforce has stated that more agents across the portfolio will adopt it over time, and that customers will eventually be able to build their own long-horizon agents with Agentforce. If you are evaluating this capability for a near-term deployment, scope it to outbound sales use cases for now.
AI Skills in Agentforce Coworker: Scaling Institutional Knowledge
Agentforce Coworker is Salesforce’s agent that works alongside employees across the surfaces where they already work — Slack, Salesforce UI, mobile. The September 11 announcement added a capability called AI Skills that addresses one of the hardest problems in enterprise AI adoption: capturing and scaling the tacit knowledge that lives in people’s heads.
The mechanism: an employee teaches Coworker how to complete a task once — say, how to process a specific type of customer escalation that involves checking three systems, applying a discount policy, and sending a templated response — and that skill becomes available to the entire workforce across all interfaces. Think of it as recording a macro, but the “macro” understands context, can adapt to variations, and works across multiple applications.
AI Skills is currently in pilot, with general availability planned for October 2026. For organizations that have struggled to codify tribal knowledge into traditional automation (a persistent challenge even for mature RPA teams), this capability is worth watching closely.
Multi-Agent Orchestration: Agents Working as a Coordinated Team
Multi-Agent Orchestration — now generally available — routes work across multiple agents so they can operate as one coordinated team when a job crosses roles, systems, or stages of the customer journey.
The coordination brain behind this is Atlas 3.0, which implements what Salesforce describes as description-driven semantic routing. Instead of hard-coding routing rules between agents, the orchestration layer reads each agent’s description of its capabilities and dynamically routes work to the agent best suited for each sub-task. This pattern — which we covered in depth in our earlier Agentforce multi-agent orchestration analysis — is becoming the dominant paradigm for multi-agent systems across platforms.
A practical example: a customer contacts Casey (the service agent) about a billing issue that turns out to involve a product return and a new purchase recommendation. Casey handles the initial service inquiry, routes the return processing to Marshall (supply chain), and hands off the product recommendation to Carter (commerce) — all within a single customer interaction, with context preserved across handoffs.
For architects designing multi-agent systems, the key question is always: what happens at the boundaries? Salesforce’s approach — grounding all agents in Customer 360 data and using a shared runtime with semantic routing — sidesteps the integration complexity that plagues multi-agent deployments built on disparate frameworks. The trade-off is vendor lock-in: this orchestration only works for agents running on Agentforce.
Agent Script: Open-Source Deterministic Control
One of the most technically interesting pieces of Salesforce’s agent infrastructure is Agent Script, an open-source, declarative language for defining agent behavior. Open-sourced at TDX 2026 in April, Agent Script addresses a problem every Agentic AI Architect has encountered: AI agents are non-deterministic, and some business logic absolutely must not be left to probabilistic reasoning.
Agent Script looks like a hybrid of YAML and Python — indentation-sensitive, with explicit variable types (including mutable boolean and mutable string for state management), declared actions, and control flow logic. The key design principle: pin the load-bearing logic that must be deterministic, and let the agent reason through the rest.
For example, a compliance rule that says “never approve a discount above 30% without manager sign-off” should not be a prompt instruction that the LLM might occasionally ignore. In Agent Script, it is a hard control flow gate that the agent cannot bypass, regardless of how it reasons about the request.
What is open-sourced: the Agent Script specification, parser, linter, compiler, language server (LSP), VS Code extension, and playground. What is not open-sourced: the runtime. Agent Script compiles to a Salesforce-internal specification format that executes on Salesforce infrastructure. You can build tooling around it, but running agents requires Salesforce’s runtime.
This is a deliberate strategic choice — open enough to build community and tooling adoption, closed enough to maintain platform stickiness. Architects evaluating this should understand the boundary clearly: Agent Script is portable as a specification, not as a runtime.
Agent Optimizer: An Agent That Improves Other Agents
Agent Optimizer, coming to GA in October 2026, is a meta-agent that works alongside teams across the agent lifecycle. It helps build and refine agents, subagents, and actions; test performance; and analyze session traces to identify what to improve.
This addresses a gap that becomes acute at scale: when you have dozens or hundreds of deployed agents, manually reviewing session logs and tuning agent behavior becomes a full-time job. Agent Optimizer automates the feedback loop — identifying patterns in agent failures, suggesting prompt or action improvements, and running performance tests.
The concept is not unique to Salesforce — IBM’s watsonx Orchestrate and Microsoft’s Agent Framework both offer observability and tuning tools — but packaging it as an agent rather than a dashboard is a bet on the agentic paradigm extending to agent management itself.
Claudeforce: Anthropic Meets Customer 360
Two weeks before the job-ready agents announcement, Salesforce and Anthropic revealed Claudeforce — a strategic partnership that marks the first time Salesforce has applied its “force” suffix to another company’s product.
The integration works in both directions. Claude moves into Salesforce as a reasoning model powering Agentforce agents. Salesforce moves into Claude as a plugin with 37 prebuilt sales skills — including meeting prep, deal health review, pipeline review, email composition, and CRM record updates — enabling sellers and agents to reason over live revenue context directly from Claude’s interface.
For practitioners, the immediate implication is model flexibility. Agentforce is not locked to a single LLM provider. Organizations can leverage Claude’s reasoning capabilities (particularly strong for complex, multi-step analysis) while keeping their data, workflows, and governance within Salesforce’s trust boundary. The open beta launched in September 2026.
The Numbers: 7 Billion Agentic Work Units and What They Mean
Salesforce reported that its platform has delivered 7 billion Agentic Work Units (AWUs) across Agentforce and Slack, including 3.2 billion in Q2 alone. An AWU represents one discrete task accomplished by an AI agent — a prompt processed, a reasoning chain completed, or a tool invoked.
These numbers are impressive as a growth signal (the 15% compound monthly growth rate suggests accelerating adoption), but architects should understand what AWUs are and are not. AWUs are a platform-level disclosure metric used in earnings reporting — there is no published price per AWU and no AWU-to-credit conversion. Enterprise billing still runs on per-user licenses (starting at $125/user/month) and per-conversation pricing ($2/conversation for the original Agentforce consumption model).
The practical takeaway: AWU volume tells you that enterprises are actually using Agentforce agents at scale, not just piloting them. Whether that usage translates to ROI depends entirely on the specific deployment — which is where the job-ready agent approach (with its pre-built skills and faster time-to-value) is designed to shift the economics.
Where Salesforce Sits in the Agentic Platform Landscape
The September 11 announcement positions Salesforce as the most vertically integrated agentic platform in the enterprise market. But how does it compare to the other major players?
| Capability | Salesforce Agentforce | Microsoft Agent Framework / Copilot Studio | UiPath Autopilot | IBM watsonx Orchestrate |
|---|---|---|---|---|
| Pre-built agents | 7 named, job-specific agents | Copilot agents across M365 apps | Autopilot coding agent + task-specific bots | Skills catalog + custom agents |
| Long-horizon execution | Yes (Hunter, expanding) | Limited (session-based) | Workflow-based persistence | Task-based persistence |
| Multi-agent orchestration | GA (Atlas 3.0, semantic routing) | GA (Azure AI Foundry) | Orchestrator-based | Skill chaining |
| Deterministic control | Agent Script (open-source spec) | Prompt Shields + guardrails | XAML workflows + coded automations | Decision engine + business rules |
| Data grounding | Customer 360 | Microsoft Graph + Dataverse | Data Service + integration connectors | Enterprise data fabric |
| Agent lifecycle tooling | Agent Optimizer (agent-as-meta-agent) | Copilot Analytics | Test Suite + monitoring | Observability dashboard |
Salesforce’s advantage is vertical integration with CRM data — if your agents need to reason over customer, sales, or service data, no other platform provides that context as natively. The disadvantage is scope: Agentforce agents are optimized for Salesforce-centric workflows. If your agentic use cases span manufacturing floors, legacy ERP systems, or desktop UI automation, platforms like UiPath and Automation Anywhere remain better suited for those domains.
What Agentic AI Architects Should Do Next
If you are evaluating Agentforce for your organization, here is a practical framework:
Start with one job-ready agent in a high-impact area. Casey (customer service) or Piper (inbound lead gen) are the lowest-risk starting points — both are GA, have proven customer results, and operate in domains where AI resolution rates are easy to measure. Avoid starting with Hunter until the long-horizon runtime moves beyond pilot.
Map your orchestration requirements early. Multi-Agent Orchestration is powerful but only works within Agentforce. If your agent ecosystem spans multiple platforms — say, Salesforce for customer-facing agents and UiPath for back-office automation — you will need an interoperability layer. The A2A protocol joining the Agentic AI Foundation alongside MCP may eventually provide that, but it is not production-ready for cross-platform agent orchestration today.
Evaluate Agent Script for compliance-heavy use cases. If your industry has hard regulatory requirements around agent decision-making (financial services, healthcare, government), Agent Script’s deterministic control flow is a genuine differentiator. The open-source toolchain means you can evaluate the language without committing to the platform.
Plan for AWU measurement from day one. Even though AWUs are not a billing unit, they are how Salesforce (and increasingly, the market) measures agent productivity. Establishing your own AWU baseline early gives you a denominator for ROI calculations that stakeholders will expect.
Watch the Claudeforce beta. If your organization uses both Salesforce and Claude, the 37-skill plugin in the September beta could materially change how your revenue teams interact with CRM data. Monitor the beta for latency, accuracy, and governance coverage before planning a rollout.
Frequently Asked Questions
What are Salesforce Agentforce job-ready agents?
Job-ready agents are pre-built AI agents designed for specific business functions — customer service (Casey), IT/HR (Paige), commerce (Carter), outbound sales (Hunter), supply chain (Marshall), inbound pipeline (Piper), and customer experience (Fin). Each ships with the skills, actions, and data models required for its job, connected to Salesforce Customer 360 data.
How does the Agentforce long-horizon runtime differ from standard AI agents?
Standard AI agents complete tasks within a single session and lose context when the interaction ends. The long-horizon runtime enables agents to pursue goals across days and weeks through persistent memory, durable execution (plans that keep running between interactions), and dynamic steering (adapting behavior based on ongoing user feedback). Currently available only for Hunter, the outbound sales agent.
What is Agent Script and is it truly open source?
Agent Script is a declarative language for defining deterministic agent behavior — ensuring critical business logic is enforced as hard control flow, not probabilistic prompt instructions. The specification, parser, linter, compiler, and VS Code extension are open-sourced on GitHub. However, the runtime that executes Agent Script is proprietary to Salesforce, so running agents requires Salesforce infrastructure.
How much does Salesforce Agentforce cost?
Agentforce pricing includes per-user licenses starting at $125/user/month and a consumption model priced at $2 per conversation. Agentic Work Units (AWUs) are a platform metric Salesforce uses for reporting agent activity, but there is no published price per AWU — enterprises are billed through standard licensing and consumption models.
Can Agentforce agents work with non-Salesforce systems?
Agentforce agents are natively grounded in Salesforce Customer 360 data. Through MCP (Model Context Protocol) integrations and the Claudeforce partnership with Anthropic, agents can access external context. However, Multi-Agent Orchestration currently only coordinates agents within the Agentforce platform — cross-platform orchestration with UiPath, Microsoft, or other agent frameworks requires additional integration work.
Key Takeaways
- Seven named agents ship with pre-built skills for service, sales, commerce, HR/IT, supply chain, and lead gen — five of seven are GA now.
- Long-horizon runtime is the architectural headline: agents that pursue goals across days and weeks with persistent memory, durable execution, and dynamic steering. Currently limited to Hunter (outbound sales).
- AI Skills in Coworker lets employees teach agents institutional knowledge once and scale it across the workforce (pilot now, GA October ’26).
- Multi-Agent Orchestration (GA) uses semantic routing via Atlas 3.0 to coordinate agents across business functions within a single customer interaction.
- Agent Script provides open-source deterministic control for compliance-critical logic, but the runtime remains proprietary.
- 7 billion AWUs delivered (3.2B in Q2 alone) signals real enterprise adoption, not just pilots.
- Claudeforce brings 37 prebuilt sales skills to Claude’s interface, with open beta in September 2026.
- Best starting point: deploy Casey or Piper in a measurable domain before evaluating Hunter’s long-horizon capabilities.
References
- Salesforce. “Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work.” September 11, 2026. salesforce.com
- Salesforce. “Salesforce and Anthropic Announce Claudeforce: The #1 AI Meets the #1 AI CRM.” August 26, 2026. salesforce.com
- Salesforce. “The Agentic Work Unit: Converting Raw Intelligence into Real Work.” 2026. salesforce.com
- Salesforce. “Agent Script: The Control Plane for Agentic Decisions.” 2026. salesforce.com
- Futurum Group. “Salesforce’s Job-Ready Agents Target Enterprise AI’s Biggest Gap.” September 2026. futurumgroup.com
- SiliconANGLE. “Salesforce introduces new AI agents to automate sales, support tasks.” September 11, 2026. siliconangle.com
- Unite.AI. “Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime.” September 2026. unite.ai
- Constellation Research. “Salesforce’s Agentic Work Unit: What you need to know.” 2026. constellationr.com
- CNBC. “Salesforce, Anthropic expand partnership as Benioff responds to ‘SaaSpocalypse’ concerns.” August 26, 2026. cnbc.com
- GitHub. “salesforce/agentscript — An open, schema-driven language for configuring agent orchestration systems.” github.com
