On September 10, 2026, OpenAI officially moved its Agents API into public beta. The core of this release is that OpenAI is directly opening to external developers the agent execution infrastructure it has long used internally—the same execution framework that powers the Codex coding agent and the ChatGPT Work product. The API is open to all developers at no cost for access, with no additional platform service fee; users pay only for model tokens, tool calls, and container runtime.
The Full Logic Behind Opening an Internal Framework
To understand this release, one must first understand why OpenAI chose this moment to open up this infrastructure. In its official documentation, the OpenAI team states clearly that in the course of running Codex and ChatGPT Work at scale, they gained deep insight into what long-running agents require—agents need an execution framework that can continuously manage context, call tools efficiently, and coordinate subagents, along with reliable infrastructure that lets them run for days on end. This framework was honed in production rather than designed from scratch specifically as a developer product.
The Agents API is built on the open-source Codex execution framework, and the entire API is organized around four core concepts: Agent (the model, instructions, toolset, and MCP server configuration), Environment (an optional sandbox for running code, reading and writing files, and loading skills), Session (a persistent agent instance that executes tasks and responds to input), and Events and Items (the task inputs sent to the agent and the outputs it produces). A session run is completed in four steps: create a session and assign a task, track progress via streaming or Webhook, and after the task is complete, continue assigning new tasks or adjust the current execution direction.
The sample code provided in the official documentation illustrates the framework's design intent: an incident investigation agent is created with a single API call, using the model gpt-6-astra, configured with up to 3 concurrent subagents responsible for deployment analysis, error analysis, and dependency analysis respectively; the main agent coordinates and summarizes the results, and finally saves the investigation conclusions, evidence, and remediation recommendations to a specified workspace path.
Three Runtime Environments, Corresponding to Three Deployment Decisions
In terms of execution environment choices, the Agents API supports three sandbox paths and can also run directly without mounting a sandbox.
The first is an OpenAI-hosted sandbox, reusing the same sandbox infrastructure behind Codex and ChatGPT; developers can configure files, packages, skills, and plugins, making it suitable for teams that want to get started quickly with zero operations overhead. The second is a self-hosted mode, in which developers run codex exec-server on their own infrastructure, register with a restricted key, and connect to the platform via WebSocket; all connections are outbound, and data does not pass through OpenAI servers. The third is partner sandboxes, which currently have first-tier integrations with nine providers—Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel—offering runtime options across different CPU, GPU, memory, and storage configurations.
This architectural design sends a clear signal: OpenAI does not intend to monopolize the agent runtime layer; instead, it is incorporating ecosystem partners' compute options into its official framework. But it also means that which sandbox path a developer chooses is in effect an architectural decision about data flow, operational responsibility, and cost structure—not merely a matter of technical preference.
Three Key Engineering Capabilities for Long-Running Execution
The Agents API includes three engineering capabilities specifically for long-running scenarios, which is what distinguishes it from ordinary stateless API calls. The first is automatic context compression: when a session approaches the token window limit, the system automatically compresses earlier information, enabling the agent to continue executing across multiple context windows without the developer needing to manually manage context truncation. The second is tool search and on-demand loading: the agent can load tool definitions on demand during execution, run multiple tool calls in parallel, chain operation results, or filter return values, reducing unnecessary context usage while preserving the model cache. The third is multi-agent coordination: developers can split complex tasks into multiple independent subtasks handled in parallel by subagents, with the main agent coordinating and summarizing; the official example configures up to 3 concurrent subagents, and the actual number is configurable.
Early User Data: Concrete Improvements in Cost and Reliability
At launch, OpenAI disclosed measured data from early-access users: the compliance technology company SafetyKit reduced its cost per case by 60% after migrating its case review workflow to the Agents API; the agent development platform Hypha cut agent response failure rates by 86% after separating its execution framework from the sandbox; the evaluation company Cirridae raised its agent score from 0.71 to 0.85, with subagent stream latency reduced to one quarter of the previous level; and the logistics company Nash.ai has used it to operate thousands of long-running agents across its global network.
The above figures all come from early partner case studies that OpenAI chose to make public, and are optimization results within their respective vertical scenarios; they cannot be directly extrapolated as expected benchmarks for general deployments.
One Clear Current Limitation: A Data Compliance Gap
The Agents API public beta has two current limitations: data is stored only within the United States, and it does not support Zero Data Retention. These are not technical details but compliance thresholds that directly affect enterprise adoption decisions.
Zero Data Retention is a standard requirement in many enterprise AI procurement contracts and one of the core selling points of OpenAI's Enterprise plan. The temporary absence of this capability in the Agents API public beta means that compliance-sensitive enterprise users in finance, healthcare, government, and the Europe and Asia-Pacific regions are effectively excluded from the compliant usable scope during the beta. For such users, the only currently viable workaround is to choose the self-hosted mode—but self-hosting means the enterprise must bear the deployment and operations costs of codex exec-server itself, which partially offsets the core selling point that "managed services reduce operational burden."
Competitive Landscape: A Comparison with the Managed Agent Platforms of Three Major Cloud Providers
OpenAI is not launching this product in a blank market. AWS Bedrock AgentCore, Azure AI Foundry Agent Service, and Google Vertex AI Agent Engine all reached general availability (GA) milestones successively between late 2025 and the first quarter of 2026, and all three cloud providers have built their own managed agent runtime products.
The managed agent platforms of the three cloud providers each have different positioning. AWS Bedrock AgentCore treats identity authentication and permission management as first-class citizens of the runtime; each agent can bind an identity via IAM or OAuth, and a secure credentials vault automatically rotates tokens—suitable for enterprises already deeply integrated into the AWS ecosystem. Google Vertex AI's advantage lies in its native deep integration with Gemini and natural connectivity to data warehouses such as BigQuery, making it suitable for GCP-native, data-intensive workloads. Azure AI Foundry's advantage lies in its fit with Microsoft's enterprise software ecosystem and its maturity in European regulatory compliance.
The OpenAI Agents API's approach differs from those three. It is not tied to any cloud platform; instead, it provides a full-stack managed service that bundles models with infrastructure, with developers using OpenAI models, the OpenAI-maintained Codex framework, and sandbox environments provided by OpenAI or its partners. This design has a notable advantage in ease of use—for teams that have already made OpenAI models their core technology choice, the path from prototype to production is the shortest. But it also means core dependencies are highly concentrated in a single vendor, with relatively higher vendor lock-in risk.
By contrast, the managed agent services from AWS and Google lean toward a "model-agnostic" positioning, both supporting access from multiple model providers. In terms of data sovereignty, IAM permissions, and depth of integration with local compliance certifications, these are often priority considerations in enterprise selection, especially for enterprises that have already built their core business on a particular cloud platform.
Strategic Assessment: What Is Most Likely to Happen Next
The launch of the Agents API has a clear strategic logic: after accumulating extensive production experience running agents through Codex and ChatGPT Work, OpenAI is opening this infrastructure to the outside world, essentially competing for the emerging market position of the "agent application runtime layer." This move marks a structural expansion of OpenAI's business model—evolving from "providing the best language model API" to "providing the best agent runtime platform." Under this framework, the model gradually becomes a component of the runtime platform, not the other way around.
There are three key signals: First, when Zero Data Retention and data localization support will appear in the general release—the arrival of these two capabilities will directly trigger a significant acceleration in the enterprise adoption curve. Second, the speed at which the ecosystem of nine partner sandboxes expands will determine actual availability and regulatory compliance coverage in non-U.S. markets. Third, how AWS, Azure, and Google respond after the Agents API launch—all three already have GA products in the market, and the next competition is likely to be rapid iteration toward feature convergence in context management cost, multi-agent coordination efficiency, and depth of compliance certification.
For developers, the more pragmatic assessment today is this: for teams in the prototype and validation stage, the OpenAI Agents API offers the shortest path on the market from idea to running a production-grade agent; once they enter the compliance assessment stage, they can then decide whether to keep, migrate, or adopt a hybrid deployment based on data sovereignty requirements and cloud platform affiliation. The pricing model during the public beta (usage-based only, with no fixed platform fee) lowers the cost of trial and error, making this a window worth exploiting. But the compliance limitations are a real threshold, not a technical detail that can be bypassed.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接