On September 10, 2026, OpenAI moved the Agents API into public beta, opening it to all developers. The API packages the agent execution layer previously used only for Codex and ChatGPT for Work into a managed service accessible through a single API call, with OpenAI responsible for running infrastructure such as session management, context compression, tool scheduling, and multi-agent coordination.
OpenAI's official announcement noted that this product path stems from experience running Codex and ChatGPT for Work at scale: long-running agents need to manage context, call tools efficiently, coordinate sub-agents, and maintain stable execution when tasks last for days. The Agents API turns these capabilities into an API.
One Call, Four Concepts
The official documentation organizes the Agents API around four core concepts: Agent (a collection of models, instructions, tools, and MCP servers), Environment (an optional sandbox in which the agent accesses files and executes commands), Session (a persistent agent instance that maintains state across tasks), and Events and Items (inputs sent to the agent and the outputs it produces).
A session lifecycle has four steps: create and assign a task, track progress via streaming or webhooks, receive output, and continue to the next task or adjust the current direction of execution. The official example shows creating a troubleshooting agent with a single call; the agent can dispatch up to three sub-agents at once, handling three parallel paths—log analysis, error tracing, and dependency checking—and ultimately writing findings and recommendations into the workspace.
This single-entry design delegates task routing, intermediate result storage, failure retries, and multi-turn session management to OpenAI infrastructure. Core issues such as permission design, acceptance testing, and decision boundaries remain the responsibility of the integrator.
Three Paths for the Runtime Environment
Developers face three sandbox choices when deploying. The first is OpenAI's self-hosted sandbox, which uses the same sandbox infrastructure as Codex and can be configured with files, packages, skills, and plugins. The second is self-hosted mode, in which developers run codex exec-server in their own environment, register with a restricted key, and establish an outbound connection via WebSocket. The third is partner sandboxes; partners with first-tier integrations include Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
The API also supports a no-sandbox mode, for scenarios that require only model orchestration and do not involve file system operations.
The Real Cost: More Complex Than "No API Fee"
OpenAI emphasizes that the Agents API itself has no separate charge, but the actual billing structure includes container hosting and tool-call fees. According to analysis by Verdent AI, container hosting is calculated per 20-minute session: $0.03 for a 1GB container, $0.12 for 4GB, $0.48 for 16GB, and $1.92 for 64GB. For a one-hour medium-sized task (4GB container), container cost alone is about $0.36, excluding model token consumption.
Tool calls are billed separately: web search costs $10 per thousand calls, and model tokens involved in the search content are added on top. Combined with the standard token rates of the underlying model (such as gpt-6-astra), the total cost of a complex multi-agent investigation task depends on task complexity, container specifications, and tool-call frequency.
Competitive Landscape: OpenAI Is Not the First
In April 2026, AWS launched Bedrock Managed Agents (powered by OpenAI) in limited preview, and on June 17 brought Bedrock AgentCore to general availability. AgentCore provides stateful sessions, each running in an isolated microVM with a file system and shell, and supports retaining files and memory across sessions.
Google's AgentSpace uses a model-agnostic, governance-oriented Kubernetes-like control plane design. Alibaba Cloud also announced a managed agent runtime layer during the same period. Managed agent infrastructure from AWS, Google, and Alibaba was all released in a concentrated wave between April and July 2026.
OpenAI's launch this time productizes its internal execution-layer infrastructure and closes the time gap relative to AWS. The first batch of integrations from nine partners, including Blaxel and Cloudflare, looks more like a signal of ecosystem positioning.
Two Hard Limits Cannot Be Ignored
The Agents API currently has two enterprise-grade limitations: US-only data residency and no support for Zero Data Retention. These two limitations reflect the data-sovereignty tension inherent in the managed infrastructure model. Self-hosted mode requires developers to maintain codex exec-server themselves, giving up the core value proposition of OpenAI handling the infrastructure.
Independent Judgment
The release of the Agents API is materially significant for two types of users: small and midsize development teams that previously had to implement session persistence, failure recovery, and multi-agent scheduling themselves; and enterprise users already using Codex who can plug the same execution layer directly into their products.
Cost depends on task complexity, container specifications, and tool-call frequency; "no API fee" cannot be used as the basis for judgment. The design of permission boundaries, output validation, and human-machine handoff points remains a problem that integrators must explicitly solve.
The starting point for evaluating this API is to list the orchestration code currently maintained manually, identify responsibilities that can be outsourced, and then weigh them against integration costs, data-compliance constraints, and the degree of dependence on a single vendor. The Agents API is essentially a business trade-off: how much control to exchange for how much engineering efficiency.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接