What Salesforce Shipped

On 11 September 2026, Salesforce expanded Agentforce with seven AI agents that have names. Not product codes. Names. Casey, Paige, Carter, Hunter, Marshall, Piper and Fin.

Each one has a job rather than a feature list. Casey handles customer service across voice, SMS, WhatsApp and web chat. Paige resolves IT and HR requests where employees already work. Carter helps shoppers compare and buy. Marshall runs back-office processes and keeps an audit record of every action. Piper qualifies inbound leads. Fin handles customer experience workflows. Hunter works a sales pipeline from research to outreach, alongside a human seller, over weeks and months.

Six are generally available. Hunter is the one still in pilot.

I want to be careful about what I am claiming here, because there is a cheap version of this post and it is not the one worth writing. Salesforce did not read my framework. A company that size arrived at this from its own customers and its own problems. That is exactly what makes it interesting.

The Convergence Is the Story

When two people solve the same problem separately and end up with the same shape, the shape is probably in the problem rather than in either of them.

I have been running a company of one with named AI Characters since March. The framework behind it says agents need an identity, a role defined by what their work produces, a written record everything flows back into, and a boundary on what they may do unsupervised. Salesforce has now shipped, at enterprise scale and enterprise expense, a product that has all four.

Their agents have names. Their agents have jobs. Their memory, in their words, preserves context and progress across sessions so work does not stop when the interaction ends. Marshall keeps an audit record of every action. Agent Script gives deterministic rules over how an agent decides, and the agents operate inside the customer's existing business rules, permissions and security rather than inside a promise.

That last one is the one I would underline. A limit that lives in the platform is a limit. A limit that lives in a prompt is a hope. Salesforce is selling the first kind, which tells you they hit the same wall everyone hits once agents get hands.

Identity, role, memory, and an enforced boundary. Two very different operations, six months apart, at wildly different budgets, landed on the same four things. That is not a trend. That is the shape of the problem.

One of Them Is Called Hunter

I had to stop and read this one twice.

REALM has five Character Classes, and one of them is the Hunter: the one who finds. It leaves the Realm, goes out into the world, and comes back with what it saw and where it saw it. Research first, then report, and the report is raw material with a source attached rather than a conclusion.

Salesforce's Hunter works a pipeline from research to outreach.

Same name. Same half of the job. Nobody copied anybody — we both reached for the obvious word for the role that goes out and finds things. That is what convergent evolution looks like when the thing being named is a role rather than a feature. You do not call it the Outbound Prospecting Module when you are trying to describe what it is for. You call it a Hunter, because everyone immediately knows what a Hunter does.

The fantasy vocabulary I use is not decoration, and this is the clearest evidence I have been given that it is not. A metaphor is how you hold a role in your head well enough to reason about it. Salesforce spent real money discovering the same thing.

What You Cannot Buy

Now the part that matters if you are one person rather than an enterprise.

Salesforce is selling you a cast. Seven agents, mapped to seven functions that exist in a company with seven departments: sales, service, commerce, IT and HR, supply chain, marketing, customer experience. If your company has those departments, that is a genuinely strong offer and you should look at it seriously.

A company of one does not have those departments. It has you, doing all of it, in an order nobody would design on purpose. There is no outbound sales function to hand to Hunter. There is a Tuesday where you need a supplier found, a listing written, and the thing pushed live, and it is all the same afternoon and all the same person.

That is the difference between a cast and a Class. A cast is a list of roles somebody else decided you have. A Class is a contract about output, and you find it by asking what the work leaves behind. A change in the world and the record of it is Warrior work. A reading of what the numbers mean is Mage work. A finding from outside with its source attached is Hunter work. Ask that question once per kind of output rather than once per agent, and one Character can hold three Classes without any rule being bent.

You cannot buy that mapping, because it is a description of your work and nobody else has seen your work.

Seven named agents is an answer to the question “who do I hire”. It is not an answer to the question “what does my work actually produce, and where does it go next”. Only the second one survives contact with a company that has no departments.

The One Still in Pilot

Six of the seven are generally available. Hunter is in pilot, with general availability promised for November, and it is the one built to pursue a goal over weeks instead of finishing a conversation.

I do not think that ordering is an accident, and I would not read it as a criticism either. Long-horizon work is simply the hard part. An agent that answers a question is bounded by the question. An agent that carries a goal for six weeks accumulates state, drifts, and makes small decisions nobody watched — and the cost of being wrong compounds quietly the whole time.

This is the same wall I hit, and it is why REALM sizes work the way it does. A Task is one sitting and one review. A Quest is a goal with an outcome you can point at. A Saga is a chapter. The hierarchy exists because “pursue this goal for six weeks” is not a unit of work you can inspect, and anything you cannot inspect you are not really supervising.

The other half of the answer is that autonomy gets cut on reversibility rather than on trust. Can the action be undone, and who sees it first? Drafting something you will read before anyone else does needs no permission. Sending that same thing to a customer does, because neither answer is comfortable. Six weeks of unsupervised outreach crosses that line constantly, which is precisely why it is the one still in pilot.

What I Would Actually Take From This

If you are running agents and have been unsure whether giving them names and roles is serious or whimsical, that question is now settled. The largest CRM company in the world just shipped seven of them with job titles and months of memory. Name your agents. Give them a role defined by what the work produces. Write everything back to one place you trust. Put the limits where the platform can hold them.

And if you are one person, do not wait to be sold a cast. The seven roles Salesforce chose are the right ones for the company Salesforce sells to. Yours are whatever falls out when you ask what your work leaves behind — and that list is yours to write, in a folder of markdown files that no vendor can deprecate.

A vendor can give you agents with names. Nobody can give you the map of your own work. That part was always going to be yours, and it is the part that keeps working when you change every tool underneath it.