Salesforce shipped seven named Agentforce agents this month. Casey, Paige, Carter, Piper, Fin and Marshall are generally available, and Hunter is in pilot with general availability expected in November 2026.
Most of the coverage reads like a feature list. That is the wrong way to look at it. Every one of these agents is a test of something you already own, and the test is not about AI at all. It is about whether the underlying record, catalogue or process is in a state that an agent can act on.
Here is what each one actually checks.
Casey checks your knowledge base
Casey handles service across voice, SMS, WhatsApp and web chat. It answers from knowledge articles. That means Casey is only as good as the last time someone reviewed those articles.
If your knowledge base has three articles on the same refund policy written in 2023, 2024 and 2026, Casey will answer confidently from whichever one it grounds on. The fix is not prompt tuning. It is a content audit and an owner for every article.
Paige checks whether IT and HR agree on anything
Paige resolves internal IT and HR requests through Slack and employee portals. This is the easiest place to start, because the audience is your own staff and a wrong answer costs a ticket rather than a customer.
What it exposes is ownership. Password resets, laptop requests and leave policy usually live in three systems with no agreed source of truth. Paige forces that conversation before it forces anything else.
Carter checks your product data
Carter assists with discovery, comparison and checkout. It does not replace a storefront. It reads your catalogue.
If two SKUs have inconsistent attributes, missing dimensions or a description written by a supplier five years ago, Carter will compare them badly. Commerce teams tend to discover their catalogue debt the week they turn this on.
Piper checks whether you have a qualification definition
Piper engages and qualifies inbound leads on the website and over email. To qualify a lead, something has to define what qualified means.
In most mid-market orgs that definition lives in a sales leader's head, and the CRM fields that are supposed to carry it are half empty. Piper applies whatever rule you give it, thousands of times, consistently. That consistency is exactly what makes a vague rule visible.
Fin checks whether your handoffs are documented
Fin orchestrates cross-channel workflows. It is the right answer when the problem is that a process breaks between teams rather than inside one.
But an agent cannot orchestrate a handoff nobody has written down. If the step between support and billing is "Sarah usually emails them", there is nothing for Fin to run.
Marshall checks whether the process is actually defined
Marshall executes back office work: invoice handling, order processing, fulfilment exceptions. Deterministic execution is the selling point, and it is also the constraint.
Marshall runs processes. It does not design them. If your exception handling is currently a spreadsheet plus judgment, that has to become a defined path first.
Hunter checks your patience
Hunter runs outbound pipeline from research through outreach, and it is the first agent built on long horizon runtime, meaning it pursues a goal across days or weeks rather than a single session.
It is in pilot. General availability is expected in November 2026. Anything you plan around it before then is a plan around a date Salesforce has not committed to.
The licensing question people skip
Each agent sits inside its product cloud, and consumption draws on the Flex Credit allowance that comes with your edition tier. Two companies on identical editions can burn very different amounts depending on conversation volume and complexity.
That is not a number you can model from a pricing page. Get it confirmed with your account executive before it lands in a budget.
Where to actually start
Start with Paige. Internal audience, contained blast radius, and the ownership problems it surfaces are the same ones that would have broken a customer-facing rollout six months later.
Then pick the agent whose input data you trust most, not the one with the biggest theoretical return. An agent pointed at bad data produces bad answers faster than a human would.
Two platform pieces shipped alongside these agents and are worth watching: multi-agent orchestration, which is generally available now, and Agent Optimizer, expected in October 2026 for building, testing and analysing agent behaviour.
Disclosure: I work at Ashapura Softech, a certified Salesforce implementation partner. The original, longer version of this piece with the full agent-by-agent breakdown and FAQ is on our blog: Agentforce Named Agents.__
