Salesforce shipped agents that work for weeks onto a meter that bills by the action — and put the cost controls in fiscal 2028

The unit of work changed. The unit of billing did not.

Agentforce has been unusually forecastable for an agent platform, and the reason is that its unit of work had edges. A conversation begins, a conversation ends. At $2 per conversation, or a handful of 20-credit standard actions at roughly ten cents each, a support organisation with a year of ticket volume behind it could build a spreadsheet and defend it in a budget meeting.

The long-horizon runtime removes the edges. Salesforce’s own framing is that agents can now “pursue goals across days and weeks instead of completing only a task or interaction,” built on three components: memory that survives sessions, durable execution that keeps a plan alive over time, and dynamic steering that adapts as feedback arrives. That is a genuine capability advance and the most interesting engineering in the announcement.

It is also a billing unit that does not exist. There is no “goal” line on the rate card. A three-week pursuit of a single B2B prospect resolves, on the invoice, into an unknown number of 20-credit actions — research steps, enrichment calls, outreach attempts, follow-ups, re-plans after a bounce. Salesforce has not published a figure for that. Not a range, not a pilot observation, not a worked example.

The 2.5 million annual credits bundled into Agentforce 1 editions works out to 125,000 standard actions. Measured against conversations, that is a comfortable allowance. Measured against durable agents that keep working while nobody is watching, it is a number with no denominator.

This desk has been tracking the same failure across vendors all year: Cursor’s always-on cloud agent subscriptions, where goal-shaped work met a meter designed for sessions; OpenAI’s Agents API, where the harness was free and the sandbox hours underneath it were not; Sakana’s Fugu Max, where orchestration tokens billed at multiples of visible output. The pattern is consistent enough to name: when autonomy lengthens, the published price stops describing the bill.

Shipping the accelerator eighteen months before the brake

The sharper problem is sequencing, and it is visible only if you read the two 10 September announcements against the 11 September one.

The AI Control Plane is the part of the Trusted Enterprise AI Harness that does the governing: register agents, apply identity and policy, observe behaviour, manage cost. It is aimed at a real problem — VentureBeat Intelligence’s July 2026 survey found 85% of enterprises running two or more agent orchestration platforms, averaging 3.1 — and Salesforce’s ambition to govern third-party agents alongside its own is the correct scope for that problem.

Salesforce says those capabilities begin rolling out in early fiscal 2028. Salesforce’s fiscal year runs ahead of the calendar, which puts that around the start of calendar 2027.

So: agents capable of running autonomously for weeks, generally available on 11 September 2026. Cost management for those agents, beginning to roll out roughly a year and a quarter later. In between sits a per-action meter with no published consumption profile and a customer holding a Pre-Commit contract.

To be fair to Salesforce, a roadmap disclosed honestly is better than one concealed, and the company was clear that the Harness is a preview. But a governance layer announced the day before the thing it governs ships is a roadmap in the strict sense: it is not a control you have. The same structure produced a $1 government ChatGPT pilot that ended with no spend cap replacing it, and an itemised bill for a 10,000-agent swarm that only became legible after the run finished.

Hunter is the demo and the disclosure

Six agents are generally available. Hunter, the outbound sales agent, is in pilot with GA planned for November 2026 — and Hunter is the first agent deployed on the long-horizon runtime, with others to follow.

That is more informative than a staggered launch usually is. The agent that best expresses the headline capability is the one still in pilot, which means the six you can buy today are largely doing interaction-shaped work on an interaction-shaped meter. The new runtime’s real consumption behaviour is, right now, a thing Salesforce knows from pilot telemetry and buyers do not.

Which makes one question unusually cheap to ask and unusually valuable to have answered: what did Hunter’s pilot cost per qualified opportunity, in credits? Salesforce has that number. It has not published it. Asking for it before November costs an email.

Salesforce also shipped Agent Script, a syntax letting developers require that specific tasks be executed in specific ways, which the company says reduces hallucination risk. Constraining an agent’s execution path is the right instinct — and it is worth reading alongside what happened when an attacker’s agent fleet ignored its own operator’s hard exclusion list this week. Scripted paths are stronger than prompt instructions. They are still not the same as a control enforced outside the model.

The Fin close reset the customer-service comparison set

The acquisition closing months early is the quietest of the three announcements and possibly the most consequential for anyone mid-evaluation.

Fin — Intercom until May 2026 — was the strongest independent in autonomous customer service: 76% claimed average resolution, 30,000+ customers, native across live chat, email, WhatsApp, SMS, voice and Slack. Salesforce signed on 15 June guiding to Q4 FY2027 and closed on 10 September, which normally signals a clean regulatory path rather than any distress. Either way, an independent option left the board a fiscal quarter earlier than buyers were planning around.

Salesforce’s stated position is that Fin remains complementary to the customisable Agentforce platform, continues to serve its customers from within Salesforce AI Labs, and keeps working with the help desks companies already run. Those are good commitments. They are also press-release commitments about a product whose strategic reason for existing has just changed, and third-party help-desk integration is exactly the kind of thing that survives an acquisition until it becomes inconvenient.

For live evaluations, the practical effect is that Sierra versus Decagon is now a comparison between two independents and a Salesforce product, not three independents. Decagon’s own valuation trajectory was built partly on being the alternative to incumbents; that positioning got both easier to argue and harder to sustain in the same week. The Sierra and Decagon reviews track where each sits, and best AI agents tools covers the wider field including the model-agnostic orchestration layer that Warp’s agent fleets represent.

What to do before you sign

1. Pilot on PayGo. Pre-Commit and Pre-Purchase buy credits against a consumption profile nobody has published. PayGo costs more per credit and less in total when the denominator is unknown.
2. Measure actions per completed outcome. Resolved ticket, qualified lead, closed return. A long-horizon agent that works for three weeks and fails burned every credit on the way, and per-run averages will hide that.
3. Ask for Hunter’s pilot consumption data now. It is the only real long-horizon cost evidence in existence, Salesforce holds it, and November is soon.
4. Build your own credit ceiling. Per agent, per period, enforced on your side. The AI Control Plane arrives in early fiscal 2028; your budget year does not wait for it.
5. Fin customers outside the Salesforce stack: get the integration promise into the contract. At renewal, not after.
6. Separate the two purchases in your business case. The six GA agents are a conventional buy on a known meter. The long-horizon runtime is a bet on an unpriced unit. Merging them into one number produces a business case that is wrong in a direction you will not discover for two quarters.

For teams weighing agent platforms more generally, best AI agents tools covers the field, GitHub Copilot’s move to token billing is the cleanest worked example of a meter change arriving after adoption, and small business buyers should note that every price in this article assumes an enterprise edition floor of $550 per user per month before a single credit is spent.

Update, 16 September 2026 — Dreamforce moved the reasoning off the meter

Three days after this article published, Salesforce used its Dreamforce keynote to announce AIforce, which delivers the platform into Claude, Slack and other external surfaces rather than through the Salesforce UI. That materially changes the analysis above, in a direction worth stating plainly: the per-action Flex Credits meter described here only runs when Salesforce does the reasoning. When Claude reads records through Salesforce’s hosted MCP server and reasons over them itself, the work bills to Anthropic in tokens and draws on your org’s existing API allowance instead of credits.

That is not a loophole so much as an unfinished price. Salesforce’s published Flex Credits rate card already names three usage types for precisely this traffic — Headless Platform Interaction, Record Operation and Process Invocation — and lists every multiplier as TBA, unmetered today, with a commitment to 30 days’ notice before metering begins. The cost-control gap this article identified therefore has a second half: the agents that spend are live now, the control plane arrives in fiscal 2028, and a third meter is built but not yet switched on. See Salesforce moved the reasoning into Claude for the full breakdown.

Questions or corrections? Email Pick Right. Want the full list? See all news.