OpenAI launched the public beta of its Agents API on September 11, 2026, giving every developer access to the framework underlying Codex and enterprise ChatGPT. The next day, Salesforce introduced seven named Agentforce AI agents-Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin-each built for a specific business function in sales, service, commerce, IT/HR, supply chain, and customer experience. Two major releases, 24 hours apart, with almost opposite philosophies about what AI agents in production should look like. One hands you infrastructure and says go build. The other hands you a named employee and says go deploy.
That gap is worth thinking through carefully before your team commits to either path.
What the OpenAI Agents API actually is
OpenAI's Agents API exposes the managed Codex harness-session orchestration, context compaction, tool selection, and multi-agent delegation-behind a single API call. Before this, if you wanted a long-running agent that didn't collapse mid-task, you had to build that plumbing yourself: checkpoint the context, handle tool failures, recover from dropped sessions. Most teams either didn't bother or reinvented it badly.
If you have ever tried to build a long-running AI agent, you know the core challenge: getting a model to plan a task is easy. Keeping the execution state, context window, and tool calls intact three hours into an unsupervised job is where most homemade setups fail.
There are no additional fees for using the Agents API-you simply pay for the tokens and tools your agents use. That pricing structure matters: it means the cost scales directly with what your agents actually do, not with a flat seat fee layered on top of your existing API bill.
The catch is data residency. Data stays US-only, and Zero Data Retention is unsupported. For teams in regulated industries or with EU data-residency requirements, that is a hard stop-not a configuration option. Highly regulated enterprises requiring ZDR or EU data residency should wait.
Client results cited by OpenAI include a 60% cost reduction for SafetyKit and an 86% decrease in failure rate for Hypha. Both figures come from OpenAI's own announcement, so read them as evidence of what the infrastructure can do, not a guarantee for your specific workload.
Beagle in action#engineering, 3:22pmWhat Salesforce is betting on instead
You are no longer being asked to buy a general assistant and work out what it can do later. That sentence from Salesforce's positioning is the clearest articulation of the competing philosophy. Instead of building every agent from scratch, companies can now start with agents designed around the jobs where AI can have an immediate business impact. Each comes with the skills, actions, and data models required for the job, and can be tailored to how each company works.
Casey handles customer service across voice, SMS, WhatsApp, and web chat. Paige takes on employee IT and HR requests through Slack and internal portals. Carter helps online shoppers compare products and checks them out in chat. Marshall runs back-office supply chain work. Piper works on websites and inboxes to qualify inbound sales leads. Fin, built on the customer experience platform Salesforce picked up through its Intercom acquisition, handles more complex support workflows across channels.
Six of the seven agents are generally available right now, while the outbound sales agent Hunter is in pilot and is scheduled to reach general availability in November 2026.
Salesforce did not disclose pricing for any of the agents or the supporting capabilities. That last sentence deserves to sit alone for a moment. You cannot run a cost-per-outcome calculation until pricing lands, and it has not.
The early customer numbers are genuinely striking. Autism Queensland reports that Paige resolves 70% of administrative requests, and Hibbett says its Hibbett AI handles 90% of core shopper journeys after going live in six weeks. Asana's Piper agent is driving four times the conversation volume, with customers deploying Piper in 45 days on average. Anthropic says Fin resolves 79% of the conversations it sees without a person stepping in.
They are also Salesforce-reported customer results, not independent proof. You should read them as early evidence, not as a guarantee that every company can copy the same numbers after a quick rollout.
Over the past two years, Salesforce has delivered 7 billion Agentic Work Units across Agentforce and Slack, including 3.2 billion in Q2 alone. That volume metric is the one number in the announcement that is hardest to dismiss. 3.2 billion discrete tasks in a single quarter is not a pilot.
The comparison that nobody is making
Two releases, one week. Same category claim-"production AI agents for enterprise teams"-but structurally different products.
| Dimension | OpenAI Agents API | Salesforce Agentforce (named agents) |
|---|---|---|
| What you get | Managed harness, you build the agent | Pre-built agents with defined roles |
| Data residency | US only, no ZDR | Inherits existing Salesforce setup |
| Pricing model | Tokens + tools, no API surcharge | Undisclosed as of September 11 |
| Time to first agent | One API call + your own tooling | Days to weeks (pre-wired skills) |
| Flexibility | High-any workflow you can describe | Bounded by the named agent's job |
| Best fit | Teams with engineering bandwidth | Teams that need to move without building |
| MCP support | Native | Via Salesforce platform connectors |
| Human approval loop | You build it | Built into multi-agent orchestration |
The non-obvious insight is that these are not competitors in the way a search result roundup would frame them. The OpenAI Agents API is infrastructure; the Salesforce agents are applications running on a different infrastructure. A team that buys Paige is not choosing between Paige and the Agents API-they are choosing between building a custom Slack-native HR agent on the Agents API versus buying a pre-built one that runs on Salesforce's platform.
The meaningful choice for most teams is about where AI-completed work actually lives. If your team's work lives in Salesforce-pipeline in CRM, support in Service Cloud, HR in a Salesforce portal-Paige and Piper are a faster path than building from scratch. Paige resolves requests across Slack, portals, and the tools employees already use. If your work is spread across a custom stack, or you are building something that does not map to a named job title, the Agents API lets developers create and operate cloud agents using OpenAI's managed Codex harness without having to build their own orchestration, context-management, and sandbox layers.
What changed this week is not that agents became possible. Teams have been running agents in production for a year. What changed is that the infrastructure cost of doing it correctly dropped significantly-and the alternative of buying a named role with pre-built skills became concrete enough to evaluate. Both shifts happened in the same 48 hours.
OpenAI Agents API for enterprise teams: common questions
What is the OpenAI Agents API and how does it differ from the existing API?
The OpenAI Agents API gives every developer the managed Codex harness as a plain API: OpenAI runs the agent loop on its own infrastructure-coordinating model calls, tool use, and context-while you supply the tools, pick the execution environment, and pay only for the tokens and tools your agents consume, with no additional fee for the API itself. The existing completions API returns a single response; the Agents API maintains sessions across turns and coordinates subagents.
Can regulated teams use the OpenAI Agents API?
Not yet, if they require Zero Data Retention or EU data residency. Data stays US-only, and Zero Data Retention is unsupported. Teams in financial services, healthcare, or with EU customers should check with legal before building production workflows on the current beta.
How do the Salesforce Agentforce named agents work in Slack?
Salesforce's named Agentforce agents sit on the existing Customer 360 data platform and operate within a company's existing business rules, permissions, and security setup. Paige in particular is specifically wired to resolve employee requests through Slack and portals. They are not standalone Slack apps-they inherit the Salesforce permission and data model.
What does an Agentic Work Unit actually mean?
Salesforce introduced the AWU in a February 25, 2026 article, defining one AWU as one discrete task accomplished by an AI agent and describing it as a platform-level metric spanning Agentforce and Slack AI. Think of it as a billable unit of agent-completed work-closer to a completed ticket or a resolved inquiry than a message or a model call.
Is it too early for most teams to commit to either platform?
For OpenAI: the API is live, free to try, and the main known blocker is data residency. For Salesforce: six agents are GA but pricing is undisclosed, and the numbers cited so far are vendor-reported. The signal from practitioners is to stop chasing more agents and start using a few well-defined ones with tight limits, human review, and clear proof-the real win is faster work on repetitive tasks without handing over risky decisions.