Salesforce Agentforce AI Agents Get Job Titles, Persistent Memory, and a Harder Test

Salesforce introduced seven named AI agents, but the defining change is not their human-friendly branding. One agent is designed to pursue sales goals for weeks or months.

The Salesforce Agentforce AI agents arrive with job descriptions spanning service, sales, commerce, employee support, and supply-chain operations. Six are generally available, while Hunter, the outbound sales agent, remains in pilot before a planned November 2026 release.

That timing matters because Hunter carries Salesforce’s largest technical promise. Its new long-horizon runtime is designed to preserve context, maintain plans, and adjust ongoing work across multiple sessions. Microsoft already offers autonomous sales agents inside Dynamics 365, so Salesforce must prove that longer memory creates better outcomes, not longer chains of hidden mistakes.

This is a shift from selling AI as an assistant toward packaging it as labor. Names and job titles make each product easier to understand, assign, measure, and place inside an operating plan. They also sharpen difficult questions about authority, supervision, and accountability.

Salesforce Agentforce AI Agents Now Arrive Ready for Specific Jobs

Salesforce is replacing the blank agent-building canvas with a catalog of agents tied to recognizable business roles.

The company announced Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin on September 11, 2026. Each agent starts with a defined area of responsibility rather than a generic chat interface.

Casey handles customer-service requests across voice, SMS, WhatsApp, and web chat. Its prepared workflows include frequently asked questions, returns, account management, and escalation to human representatives.

Paige focuses on employee support. It addresses IT and human-resources requests through Slack, company portals, and other workplace systems.

Carter serves shoppers. It can help customers find and compare products, answer questions, and complete purchases within a conversation.

Marshall is aimed at supply-chain and back-office processes. Salesforce says it combines AI reasoning with deterministic execution, meaning fixed rules control actions that should not depend on model judgment. It also records each action for later review.

Piper qualifies inbound business leads across websites and inboxes. Fin handles customer-experience workflows across messaging, email, phone, and Slack.

Hunter is the outlier. According to the official agent portfolio, it researches prospects, conducts outreach, and works with sellers across weeks or months. Salesforce plans to make it generally available in November.

Customers can change the agents’ names and adapt their behavior. That customization turns the names into templates rather than permanent product identities.

The job titles still serve an important commercial function. A buyer can compare “help agent” performance with a support team’s existing metrics. A sales leader can evaluate an outbound agent against pipeline targets. A supply-chain manager can examine whether an automated process followed required approvals.

A generic assistant lacks that obvious unit of accountability. It can answer questions across many topics without owning a measurable outcome. These new agents are presented as owning bounded work.

That positioning also reduces the effort required to imagine a use case. Companies no longer need to begin by asking what an agent might do. They can start by evaluating whether a prepared role matches an existing process.

Salesforce says its seven agents include the necessary skills, actions, and data models for their assigned jobs. Buyers must still connect company data, configure permissions, define business rules, test edge cases, and choose escalation points.

“Ready for the job” therefore does not mean ready to operate without preparation. It means Salesforce has moved more design decisions into the product before deployment.

This approach makes enterprise AI easier to buy. It also makes failure easier to attribute. If Casey mishandles returns or Hunter pursues unsuitable accounts, buyers know which agent and workflow require investigation.

The launch creates the article’s central tension: Salesforce is making agents look more like dependable colleagues just as their work becomes more persistent and difficult to supervise.

Long-Term Memory Turns Hunter Into the Real Test

Hunter matters because it tries to maintain a business objective after the conversation that started it has ended.

Most workplace chatbots operate within a request-and-response loop. A person asks for information, receives an answer, and decides what happens next.

Hunter is designed for work with a longer time horizon. A seller might ask it to identify at-risk deals before a quarter closes. The agent could then convert that objective into tasks, locate relevant information, conduct outreach, and update its plan.

Salesforce attributes this behavior to three connected capabilities: memory, durable execution, and dynamic steering.

Memory preserves context and progress between sessions. Durable execution keeps a plan active and lets the agent resume after interruptions. Dynamic steering changes its behavior when a user supplies new information or direction.

Persistent memory is different from a large context window. A context window is the information a model can process during one inference. Persistent memory stores selected state so work can continue across separate interactions.

Durable execution is equally important. The agent needs a record of completed steps, pending actions, dependencies, failures, and approval requirements. Otherwise, remembering a goal does little to ensure reliable progress.

Dynamic steering introduces another layer. A salesperson may change priorities, remove an account, revise messaging, or require approval before further outreach. Hunter must incorporate those instructions without losing earlier constraints.

That combination resembles a workflow engine with an AI planning layer. The model proposes and adapts actions, while the surrounding system maintains state and enforces boundaries.

It also creates a new failure pattern. A wrong answer in a chatbot is usually visible immediately. A wrong assumption inside a month-long plan can quietly influence many later actions.

For example, Hunter might classify an account incorrectly, select an unsuitable contact, or interpret silence as permission to continue outreach. Each later task can remain internally consistent while still serving a mistaken premise.

That risk makes traceability essential. Buyers need to see what the agent believed, which information it used, what changed its plan, and when a human approved an action.

Salesforce says Hunter can distinguish between actions it may take autonomously and actions requiring seller approval. The company has not published independent evidence showing how reliably those boundaries hold across varied deployments.

Long-lived memory also needs active maintenance. Stored facts can become outdated, duplicated, or detached from their original context. A useful AI knowledge base needs provenance and retrieval controls, not merely more stored text.

The same principle applies to Hunter. Remembering months of activity is valuable only when the agent retrieves the correct history at the correct time.

Salesforce plans to extend the long-horizon runtime to other agents and eventually let customers build their own long-horizon agents. Hunter is therefore both a product and a production test for the underlying architecture.

If Hunter works as described, Agentforce can move beyond isolated assistance into continuing operations. If it requires constant correction, the longer time horizon will amplify supervision costs.

The Pressure Falls on Microsoft, ServiceNow, and Enterprise Buyers

Salesforce is not creating the enterprise-agent market, but it is forcing competitors to explain who owns an outcome over time.

Microsoft already provides specialized agents for sales qualification, opportunity research, closing, and recommended actions. Its sales agent catalog includes agents that research leads, send outreach, identify opportunity risks, and assist throughout a sales cycle.

That makes Microsoft the clearest competitive reference for Hunter. Both companies can ground agents in customer records and workplace communications. Both are turning systems of record into systems that initiate work.

Microsoft’s advantage comes from its reach across Outlook, Teams, Microsoft 365 Copilot, Power Platform, and Dynamics 365. Signals from meetings, email, documents, and CRM records can support sales decisions within tools employees already use.

Salesforce’s argument centers on Customer 360, business processes, and the long-horizon runtime. It wants buyers to believe that an agent embedded deeply in CRM can maintain a goal better than an assistant moving between productivity applications.

The contest is therefore not simply Salesforce versus Microsoft on model quality. It concerns orchestration, context, permissions, and the durability of state.

ServiceNow approaches the problem from workflow and governance. Its expanded AI Control Tower is designed to discover, observe, secure, and measure agents across multiple enterprise systems.

That emphasis becomes more relevant when agents remain active for weeks. A company might use Salesforce for sales, ServiceNow for IT operations, Microsoft for workplace productivity, and custom agents for internal processes.

A named agent can still cross several systems while pursuing one outcome. The buyer must decide which platform records authority, monitors behavior, and resolves conflicts between agents.

Salesforce is responding with its own orchestration and optimization features. Multi-Agent Orchestration is generally available, according to the company. It routes work among specialized agents when a job crosses roles or systems.

Agent Optimizer is scheduled for general availability in October 2026. Salesforce says it will analyze session traces, help teams test performance, and identify possible improvements.

AI Skills for Agentforce Coworker are also planned for October. They are intended to let employees demonstrate a task and reuse that knowledge across the workforce.

These capabilities reveal the actual competitive layer. Creating one impressive agent is no longer enough. Enterprise vendors need a system for teaching, coordinating, evaluating, and governing fleets of agents.

The pressure also falls on buyers. A prepared role reduces initial design work, but it can encourage companies to automate a process before examining whether that process is stable.

A badly defined sales policy does not improve because Hunter executes it more frequently. An inconsistent returns policy remains inconsistent when Casey applies it across more channels.

Companies must separate automation readiness from product availability. General availability means a vendor offers and supports a product. It does not mean every organization has the data quality or operating discipline to deploy it responsibly.

Microsoft, ServiceNow, and Salesforce now describe agents as active participants in business processes. The next stage of competition will focus less on whether agents can act and more on whether organizations can trust the resulting work.

Named Agents Make AI Easier to Buy and Harder to Excuse

Human-style names simplify adoption, but job titles also create expectations that feature branding cannot satisfy by itself.

Software vendors have long packaged complex systems around roles. Sales platforms serve sellers, service platforms serve support teams, and commerce platforms serve merchants.

Salesforce is taking that familiar structure one step further. The product itself now receives the role.

That framing helps nontechnical executives understand the portfolio. Casey is easier to discuss than a collection of service actions, retrieval components, channel integrations, and model settings.

The name creates a mental shortcut. The job title supplies the expected output. A deployment conversation can begin with responsibilities rather than architecture.

This is useful marketing, but it can also blur the distinction between software and employment. A human colleague brings contextual judgment, informal knowledge, ethical responsibility, and the ability to challenge an unsuitable instruction.

An AI agent performs within configured systems and permissions. It lacks personal accountability, even when its interface uses a name and conversational voice.

The practical unit of responsibility remains the organization. Managers decide what data the agent can access, what actions it can take, when approval is required, and how incidents are investigated.

That distinction becomes critical for long-running work. A chatbot mistake may affect one response. An outbound sales agent can contact many prospects, modify records, and shape a pipeline before someone notices a pattern.

A named role should therefore come with a formal operating description. That document needs permitted actions, prohibited actions, escalation thresholds, data sources, retention rules, owners, and review schedules.

The agent also needs outcome metrics that account for quality. Pipeline volume alone can reward low-quality outreach. Resolution rate alone can hide reopened cases, customer frustration, or incorrect answers.

Salesforce has supplied several customer-reported results. Engine says its help agent fully resolves 50% of chat inquiries. Perk reports that Hunter builds 60% of its sales pipeline.

Autism Queensland says Paige resolves 70% of administrative requests. Hibbett reports that its shopping agent handles 90% of core shopping journeys after a six-week deployment.

Asana says Piper drives four times its previous conversation volume, while customers deploy Piper in 45 days on average. Salesforce also says Fin autonomously resolves 79% of the Anthropic conversations it handles.

These figures show why packaged agents attract buyers. They connect an agent directly to operational outcomes instead of model benchmarks.

However, they are customer examples selected and published by Salesforce. They do not establish typical performance, and their denominators may differ across businesses.

“Resolved” can mean different things depending on escalation policies, case definitions, and measurement windows. “Pipeline built” does not necessarily mean revenue closed.

The useful response is not to dismiss the results. Buyers should ask for precise definitions, baseline comparisons, exception rates, and evidence from deployments resembling their own.

Names help people remember products. Job titles help leaders assign budgets. Only measured performance under real operating conditions can turn those labels into durable trust.

Salesforce’s Evidence Is Promising but Still Vendor-Reported

The strongest argument for Agentforce comes from specific usage data, yet most of the launch evidence remains controlled by Salesforce and its customers.

Salesforce says it has delivered 7 billion Agentic Work Units across Agentforce and Slack. Agentic Work Units are the company’s consumption measure for agent actions and related work.

Of that total, 3.2 billion units were delivered during the company’s second fiscal quarter. The number indicates substantial activity, but it does not reveal how many tasks succeeded or required correction.

Salesforce’s August 26 quarterly results reported that Agentforce annual recurring revenue grew more than 240% year over year. Agentic Work Units in the quarter grew 97% from the previous quarter.

Those figures show that Agentforce is becoming commercially important to Salesforce. They also explain why the company is packaging agents around recognizable jobs.

Usage growth alone cannot validate autonomy. One customer might generate many low-risk actions, while another produces fewer actions with serious financial or reputational consequences.

The seven agents also occupy different risk categories. Carter answering a product question differs from Marshall changing a supply-chain process. Hunter sending external messages differs from Paige retrieving an internal policy.

Each role needs its own evaluation design. An overall Agentforce success rate would conceal the operational differences.

Fin provides Salesforce with a more mature customer-service product and an installed base. The company completed its Fin acquisition one day before announcing the expanded agent portfolio.

Salesforce says Fin serves more than 30,000 companies and averages a 76% resolution rate. Fin continues operating within Salesforce AI Labs while joining the broader Agentforce portfolio.

That history makes Fin different from a newly introduced template. It brings an existing product, specialized customer-experience models, deployment knowledge, and customer relationships.

It also exposes a strategic reality. Salesforce is combining internal development with acquired products to accelerate its agent portfolio. Piper similarly reflects Salesforce’s earlier acquisition of Qualified.

This combination can shorten the path to useful agents. It can also create integration work across data models, administration tools, evaluation methods, and product identities.

Customers should watch whether Fin, Piper, and Salesforce-built agents eventually share consistent controls. A portfolio is more valuable when administrators can apply common policies and inspect comparable performance.

Memory raises separate questions about privacy and data governance. Long-running agents must retain enough information to continue a task, but they should not preserve every detail indefinitely.

Enterprises need controls for what enters memory, how long it remains, who can inspect it, and whether users can correct or delete it. Salesforce’s launch explains memory as a capability but does not publish detailed answers for every deployment scenario.

Security teams will also want to test indirect prompt injection. A sales agent may encounter malicious instructions inside emails, web pages, documents, or CRM fields.

An agent that can act across weeks gives attackers more opportunities to influence a plan. Memory could preserve a hostile instruction after the original content disappears from immediate view.

These risks do not make long-horizon agents impractical. They change the standard of proof.

A successful pilot should measure more than completion. It should examine authorization errors, stale-memory retrieval, incorrect plan changes, unnecessary outreach, human override rates, and recovery after interruptions.

Salesforce has given buyers concrete products to test. The company has not yet shown that a named agent can perform months of work with employee-like reliability.

Three Signals Will Show Whether Long-Horizon Agents Are Ready

The next test is not another polished demonstration. It is whether Hunter survives production use with measurable control and acceptable supervision costs.

The first signal is Hunter’s planned November 2026 general availability. Salesforce must show that the long-horizon runtime can leave pilot status without narrowing its central promise.

Availability details will matter. Buyers should examine which actions Hunter can perform, which require approval, and whether administrators can inspect changes to a plan over time.

Deployment documentation should explain how memory is selected, updated, and removed. It should also show how an interrupted task resumes without repeating work or ignoring newer instructions.

If Hunter reaches general availability with clear controls, the launch thesis becomes stronger. A delay, narrower scope, or heavy approval requirement would suggest that multiweek autonomy remains difficult.

The second signal is evidence from production deployments. Salesforce’s current examples are useful, but buyers need broader reporting on quality and intervention.

For Hunter, the key measures include qualified pipeline, response quality, unwanted-contact rates, human correction, and conversion beyond initial engagement. Raw outreach volume would reveal little.

For Casey and Fin, resolution should be paired with reopened cases, escalations, customer satisfaction, and verified answer accuracy. For Paige, organizations should track incorrect policy guidance and unnecessary access to employee data.

For Marshall, auditability deserves special attention. Buyers need to know whether deterministic controls can reliably contain AI-generated plans when supply-chain conditions change.

Independent case studies would strengthen Salesforce’s claims. Consistent results across industries would matter more than another set of exceptional customer examples.

The third signal is the competitive response from Microsoft and ServiceNow. Microsoft already describes Dynamics 365 Sales as a system of action with autonomous research, qualification, outreach, and closing agents.

If Microsoft emphasizes durable plans across weeks, Salesforce’s runtime will quickly become a category requirement rather than a lasting distinction. If Microsoft focuses on unified Microsoft 365 context, the competition may center on which data environment best supports a sales agent.

ServiceNow’s response will likely focus on governance and cross-platform observation. That approach becomes more valuable as companies deploy agents from several vendors.

A buyer may ultimately choose Salesforce for customer workflows while using another platform to oversee a broader agent estate. Salesforce must show that its orchestration and optimization tools can function within that mixed environment.

The larger question is whether an agent becomes more valuable when it acts longer or merely more difficult to monitor. Names and job titles cannot answer that question.

Salesforce has made a clear bet. Enterprise AI will be adopted through defined roles, prepared skills, persistent memory, and measurable business outcomes.

That bet places the company beyond the chatbot era. It also raises the cost of failure because these agents are expected to carry responsibilities, not simply produce suggestions.

For enterprise buyers, the sensible next step is a bounded production test with explicit controls. Select one process, document its decision rights, and measure both outcomes and correction costs.

Developers should focus on state management, observability, evaluation, and recovery paths. Knowledge workers should ask which judgments remain theirs and whether the system exposes enough context to challenge its decisions.

The Salesforce Agentforce AI agents are easier to understand because they now resemble a team. The decisive evidence will come when organizations can show that these named roles remain accurate, controllable, and worth supervising over time.