Salesforce Agentforce Adds 7 Named Agents and a Weeks-Long Runtime

TL;DR

Salesforce announced 7 named Agentforce agents on 11 September: Casey, Paige, Carter, Hunter, Marshall, Piper and Fin. 6 are generally available; Hunter is in pilot with GA in November 2026.

Hunter runs on a new long-horizon runtime that pursues goals "across days and weeks using memory, durable execution, and dynamic steering", rather than inside one chat session.

Salesforce reports 7 billion Agentic Work Units delivered, 3.2 billion in Q2 alone. That is 46% of everything it has ever counted, in one quarter.

Three days ago a survey found 76% of organisations believe they can stop an agent in 15 minutes and 33% have a kill switch. The goal horizon just went from minutes to weeks.

Microsoft names agents by job function. Salesforce names them like colleagues. That is a product decision with governance consequences.

What shipped, and what each one does

Seven agents, aimed at named jobs rather than at capabilities.

Agent. Job. Availability

Casey. Help agent. Resolves service issues across voice, SMS, WhatsApp and web chat. GA now

Paige. IT and HR service agent. Employee requests across Slack, portals and workplace tools. GA now

Carter. Shopper agent. Product discovery, comparisons and checkout. GA now

Marshall. Supply chain agent. Back-office orchestration and manual workflow automation. GA now

Piper. Inbound pipe-gen agent. Engages, qualifies and converts inbound leads. GA now

Fin. Customer agent. Complex customer-experience workflows across channels. GA now

Hunter. Outbound sales agent. Pipelines from research through outreach. Pilot, GA November 2026

Alongside them: Multi-Agent Orchestration, routing work across specialised agents, GA now; and Agent Optimizer, for building, refining, testing and analysing agent performance, GA October 2026.

The number that describes the curve

Salesforce published a usage figure that is more informative than the product list.

7 billion Agentic Work Units delivered in total, of which 3.2 billion landed in Q2 alone.

2 / 7 = 45.7%

Nearly half of all agentic work Salesforce has ever counted happened in a single quarter. Whatever an Agentic Work Unit is exactly, and the company defines it rather than an auditor, the ratio is the point. This is not a curve flattening out. It is a curve where the most recent quarter is comparable to every prior quarter combined.

Why the runtime matters more than the seven agents

Because it changes the unit of risk, and almost nothing in enterprise governance is built for the new one.

Hunter is the first agent on a long-horizon runtime that pursues goals "across days and weeks" with memory, durable execution and dynamic steering. Every previous generation of this product worked inside a conversation. A session began, something happened, the session ended.

Run the change as a ratio. Take a generous chat session at 30 minutes against a three-week goal:

chat session 30 minutes

three weeks 30,240 minutes

about 1,000x longer

A control designed for a thirty-minute session now has to hold for three weeks. Those are not the same engineering problem. A session that goes wrong is bounded by the session. A durable goal that goes wrong is bounded by whoever notices, and by whether the thing that notices is still running.

Three days ago we covered two surveys measuring the gap between what companies believe about their agents and what they can do. The worst pair was the kill switch: 76% believe they can disable an agent within 15 minutes, 33% have the mechanism. That gap was already uncomfortable when agents lived in chat sessions. Against a goal horizon of weeks, a fifteen-minute stopping assumption is not a control. It is a hope with a stopwatch attached.

Does naming an agent "Casey" change anything?

Yes, and it is worth being precise about how, because the effect is real and it runs both ways.

Compare the two naming conventions shipping this quarter:

Salesforce. Microsoft Business Central

Naming. Casey, Hunter, Paige. Payables Agent, Sales Order Agent, Expense Agent

Reads as. A colleague with a job title. A feature with a scope

What naming buys. Adoption. A support team that can say "Casey handled it" has a vocabulary for the work, and a name is easier to assign a queue to than a capability is. Salesforce is solving a real onboarding problem, and the functional naming convention has its own failure mode: nobody knows who owns the Payables Agent either.

What naming costs. A person-shaped name invites person-shaped assumptions, and the assumptions are wrong in exactly the places that matter. A colleague has judgment, accountability and a manager. An agent has a configuration, a permission set and a budget. When something goes wrong with Casey, the correct question is which permission and which prompt, not which decision. Names make the first question harder to ask.

Neither convention is a governance control. The register is.

What about governance?

Salesforce shipped the piece of this that most organisations measurably lack, which is the interesting part.

Its platform architecture groups context, agency, action, governance and security, with a control plane that registers agents and manages their lifecycle. A registry of what exists, what it can reach and when it was retired is precisely the inventory control the Harness survey found at 44% adoption against 77% confidence.

So the vendor is closing a real gap. With one boundary worth naming plainly: a control plane registers the agents inside its own platform. An estate running Agentforce alongside Business Central agents, a few Copilot Studio agents and whatever finance built in a weekend has a register per vendor and no register of the estate. When we compared how four vendors govern agents, that was the finding then and this does not change it. Per-platform governance is real governance for that platform and an incomplete answer to "what are we running".

What should a buyer do about this?

Ask what the stopping story is for a long-horizon agent, in writing. Not whether it can be paused mid-session. Whether a goal in flight for eleven days can be halted, what happens to work already committed, and who can do it at 3am.

Treat durable memory as a data question. An agent that carries memory across weeks is holding customer context outside the conversation that produced it. Where it lives, how long it persists and what deletes it are retention questions, not product questions.

Check the boundary of the registry before you rely on it. A control plane that covers one vendor's agents is useful and is not an inventory. Ask specifically whether it can see agents it did not create.

Start with the six, not the seven. The generally available agents run in the familiar session model. Hunter is the one that changes the risk shape, and November is enough time to decide deliberately rather than by pilot momentum.

FAQ

How many of the seven can I use today?

Six: Casey, Paige, Carter, Marshall, Piper and Fin. Hunter is in pilot, with general availability stated for November 2026.

What is the long-horizon runtime?

A runtime that lets an agent pursue a goal across days and weeks using memory, durable execution and dynamic steering, rather than completing work inside a single session.

What is an Agentic Work Unit?

Salesforce's own usage metric. It reports 7 billion delivered in total and 3.2 billion in Q2. The company defines the unit, so treat it as a growth signal rather than a comparable industry measure.

Is Multi-Agent Orchestration available?

Yes, generally available now. Agent Optimizer follows in October 2026.

Does this replace the existing Agentforce agents?

No. These are prebuilt agents for named job functions, added to the platform rather than replacing the ability to build your own.

The last mile

The headline is seven names. The change is the runtime.

Enterprise software has spent two years shipping agents that live inside a conversation, which is a shape everyone already knew how to reason about: it starts, it ends, and the blast radius is the session. A goal that persists for weeks is a different object. It accumulates state, it survives the person who started it, and it keeps making decisions on days when nobody opens the tab.

That is not an argument against it. Long-running goals are where the actual work is, and a system that forgets everything between sessions was never going to run a sales pipeline. It does mean the governance question has quietly changed from "what can this agent do" to "what is this agent still doing, and who would notice if it stopped being right". Answering that across every vendor in an estate, rather than one register at a time, is the layer Cognilium builds in, and this week it got about a thousand times more time-sensitive.