On September 15, 2026, Salesforce introduced a portfolio of job‑ready agents and a new long‑horizon runtime under its Agentforce banner, a move flagged in a weekly update by AI Agent Store. The release folds sales, service, commerce, HR/IT, and supply chain use cases into packaged agents that plug into Customer 360 and ship with pre‑built skills, actions, and data models.
What the Salesforce Agentforce launch includes
According to AI Agent Store’s September 15 brief, the bundle pairs job‑ready agents with a long‑horizon runtime that lets them pursue goals over days or weeks. Salesforce is also highlighting an open‑source Agent Script language to define behavior, general availability of Multi‑Agent Orchestration, and tooling to teach and continuously improve agents over time. The point is speed: start with working agents instead of building from scratch, and let them operate against the context already modeled in Customer 360.
That packaging matters in CRM, where much of the work is already encoded as objects, records, entitlements, and flows. Giving agents pre‑built actions and a data model reduces integration effort. It also narrows the blast radius: an agent inherits the same reference data, permissions, and record types admins already manage. The Salesforce Agentforce launch positions agents as first‑class actors inside that control plane rather than as a bolt‑on chatbot.
Why a long-horizon runtime changes enterprise work
Most assistants answer, then disappear. Long‑horizon agents persist, react to new signals, and keep working until they hit a definition of done. In sales, that can mean tracking an at‑risk opportunity for a week, nudging stakeholders, logging calls, and escalating on missed responses. In service, it might span a full return and refund flow that touches inventory, finance, and customer comms. AI Agent Store’s summary cites examples like outbound lead qualification and returns handling to start small and prove value.
Running for days introduces engineering requirements that chat‑style bots rarely meet. Agents need durable state, reliable scheduling, idempotent actions, and clear failure semantics so they retry safely instead of duplicating work. They also need observability. If an agent misses a dependency or a tool returns no results, the system should expose that clearly and guide the human next step. Those patterns are common in mature software operations—think idempotency—but only now are they being pulled into agent runtimes at scale.
The compliance angle looms large. Multi‑day autonomy without rigorous logging is a risk. Salesforce’s framing, per AI Agent Store, stresses mapping Agent Script to existing business rules and audit logs for review. That nods to a governance model enterprises already understand: put the agent inside the same policy envelope as the user who would do the work, record every action with context, and keep approvals where they belong. Frameworks like the NIST AI Risk Management Framework provide a checklist mindset here: define intended use, enumerate risks, and instrument for monitoring.
What this means for admins and developers
For admins, the promise is shorter setup because the agents are designed for Customer 360. That reduces custom connectors and bespoke data stitching. The Salesforce Agentforce launch also suggests a shift in day‑to‑day work: tuning Agent Script policies may start to look like managing validation rules or flows. The test discipline should match that change. Treat agent behavior as versioned configuration, promote it through sandboxes, and attach change requests to any policy edit that enables high‑impact actions.
For developers, the interesting piece is the long‑horizon runtime. Durable orchestration invites patterns they already use in workflow engines: compensation steps, timeouts, rate limits, and SLOs. If multi‑agent coordination is generally available, teams can split roles across specialist agents—one focused on data gathering, another on approvals, a third on updates—and keep them synchronized by design. External reading on multi‑agent coordination, even an overview like multi‑agent systems, helps frame the trade‑offs between parallelism and control.
How to measure value and keep risk in bounds
AI Agent Store recommends piloting a single, narrow, high‑value flow and tracking cycle time and hand‑off rates. That’s a sound starting line. Add a few concrete guardrails before go‑live:
- Define “done.” A long‑running agent needs an explicit stop condition and a timeout.
- Instrument every action. Pair agent activity with audit events, then sample and review them weekly. Salesforce’s Event Monitoring shows the kind of telemetry to expect.
- Gate high‑risk steps. Keep human approval for discounts above a threshold, or for any irreversible financial posting.
- Track user trust. If humans often override or halt the agent, find the pattern and tune policies or tools.
Cost control is part of the assessment. Long‑running work consumes compute and API calls over time. Set usage budgets per agent, set caps for external calls, and alert on drift. A useful early indicator is the ratio of successful completions to retries. If retries swell, the problem is usually tool quality or brittle prompts, not the runtime itself.
What to watch next from the Salesforce Agentforce launch
Three signals will show whether this shift sticks. First, depth of packaged skills. If the out‑of‑box actions cover the messy middle of real processes—returns with exceptions, escalations with SLAs—adoption goes faster. Second, ecosystem energy around Agent Script. An active open‑source spec with examples, linters, and test harnesses would lower the bar for safe customization. Third, the operational polish of long‑horizon execution: clear failure states, reliable resumability after outages, and clean recovery paths when humans intervene.
The Salesforce Agentforce launch points to an era where agents don’t just answer questions; they own outcomes over time, inside the same guardrails as the systems that pay the bills. If that promise holds up in pilots, CRM work won’t feel like chat. It will feel like auditable automation that keeps going after everyone logs off. For more on this, see reuters.com and bloomberg.com and nytimes.com.