Named Agents Became Products: Salesforce Ships Seven, Buys Fin — and a Report Puts Confidence at 77% vs Control at 44%

What happened

On September 11 Salesforce unveiled seven Agentforce agents with job titles and first names: Casey for customer service, Paige for IT and HR requests, Carter as a shopping assistant, Marshall for supply-chain back office, Piper for inbound lead qualification and Fin for customer experience — all generally available now — plus Hunter, the outbound sales agent, in pilot with GA planned for November. Alongside came a long-horizon runtime for goals that span days and weeks, Multi-Agent Orchestration (GA) and an AI Control Plane for governing agents from multiple vendors (preview now, broad availability early 2027).

A day earlier, on September 10, Salesforce announced it had completed the acquisition of Fin (formerly Intercom). According to Salesforce, Fin is a customer-agent platform used by more than 30,000 companies with an average resolution rate of 76%; terms were not disclosed. The announcement's outcome numbers are specific: 7 billion cumulative "agentic work units" across Agentforce and Slack, 3.2 billion in Q2 alone, 60% of Perk's pipeline built by Hunter, and 79% of Fin-handled conversations at Anthropic resolved autonomously.

Also on September 10, Harness published "The State of Agent DLC 2026." Surveying 700 technology leaders at large enterprises (1,000+ employees, $100M+ revenue) across the US, UK, France, Germany and India, it found a 30-to-55-point gap between confidence in agent controls and the controls actually in place.

Why it matters: the industry and economic context

This week's signal is that agents have become job-level products — SKUs. Until last year vendors sold platforms and builders; now they sell finished agents with names and job descriptions and lead with customer numbers like a catalogue. The Fin deal vertically integrates channels (chat, email, WhatsApp, voice, Slack), proprietary models and 30,000 customers into that line in one move, and landing days before Dreamforce (September 15–17) reads as a land grab for "job agents" against ServiceNow, Microsoft and Google Cloud. Add Cisco's late-August rollout of a personal agent, MyAgent, to all 90,000 employees — with agentic interactions up nearly 350% quarter over quarter — and demand is crossing from pilots to company-wide deployment too.

Two numbers, same week	Supply side (vendor claims)	Demand side (Harness survey, n=700)
Visibility	30,000 Fin customers folded in	77% confident of a complete inventory vs 44% running discovery tooling
Quality	Fin average resolution 76%	74% trust testing to catch failures vs 19% with an automated release gate
Response	Multi-Agent Orchestration GA	76% believe they can disable an agent within 15 minutes vs 33% with an instant kill switch
Cost	3.2B work units in Q2	74% claim full per-agent spend visibility vs 60% who overran budget last quarter

The left column is vendor-selected case studies; the right says most organizations lack the controls to reproduce those numbers at home. That both arrived in the same week is the story.

What it means for practitioners: decisions to make now

The principle — vendor outcome numbers are candidate metrics, not your metrics. "76% resolution" is an average over the channels and industries Fin was trained on; "60% of pipeline" is one customer. An evaluation should start by measuring the same definitions on 200–500 of your own tickets or leads, and teams handling Korean-language support or domestic payment and delivery exceptions should sample each language separately.

The common failure pattern: a pilot performs, licenses expand, and the agent inventory, change pipeline, release gate and shutdown procedure still don't exist. 37% of respondents route fewer than half their agent changes through a pipeline, and 58% report more incidents per 100 changes since deploying agents. Bringing in several named agents multiplies the problem — "orchestration is GA" also means one agent's prompt change can alter another agent's output.

The response is to stand up four controls before signing: automated discovery of agents, MCP servers and models; a regression gate that puts prompt and tool-schema changes on the same pipeline as code; a kill switch you have actually rehearsed, targeting one minute rather than fifteen; and per-agent cost in your accounting unit (cost per resolved case), not work units. The vendor's Control Plane stays in preview until early 2027, so that gap is yours to close — and Korean organizations should check the contract for cross-border transfer and vendor-training use of support data.

Checklist

Re-measured vendor case numbers on 200–500 of your own samples with the same definitions — Korean-language samples separately.
Agent, MCP-server and model inventory maintained by automated discovery (a hand-kept list puts you in the "77% confident" column).
Prompt and tool-schema changes pass the same pipeline and regression gate as code (see the 95% agreement release line).
Rehearsed shutting one agent down within a minute this week.
Converted monthly per-agent cost into cost per case and placed it next to the human cost.
Checked the contract for cross-border transfer and vendor-training use of support data.

What to watch

Over the next four weeks: whether Dreamforce (September 15–17) brings pricing and a Fin integration roadmap, whether Agent Optimizer reaches GA in October and Hunter in November as scheduled, and whether competitors answer with their own catalogues of named job agents. Whether the Control Plane really governs other vendors' agents won't be verifiable until early 2027.

References

Ask AI about this article

The assistant has read this article. Ask anything — it answers from the text and says so when something isn't in it.

Loading the chat…