Back to list
Product LaunchOpenAIAgents APICodex

OpenAI Introduces Agents API: A Managed Codex Harness for Cloud Agent Orchestration and Tool Use

OpenAI has officially announced the introduction of the Agents API, a dedicated managed service designed to help developers build and launch autonomous cloud agents. Powered directly by OpenAI's Codex harness, the new API provides managed infrastructure engineered for orchestration, persistent long-running sessions, and reliable tool use. By packaging complex runtime management into a cloud-based service, the Agents API simplifies how developers deploy agents capable of handling extended, multi-step workflows. This release marks a significant step forward in operationalizing autonomous systems, establishing harness-level orchestration as a managed platform standard.

OpenAI Blog
Share

Key Takeaways

  • Managed Agent Infrastructure: OpenAI has unveiled the Agents API, a managed cloud service that allows developers to build and launch AI agents without maintaining custom backend orchestration.
  • Codex Harness Engine: The runtime is powered by OpenAI's Codex harness, providing the underlying scaffolding necessary for reliable agent execution and session persistence.
  • Long-Running Session Support: The service is explicitly engineered to handle long-running sessions, solving one of the most critical challenges in autonomous multi-step task execution.
  • Native Tool Integration: Built-in orchestration and tool use capabilities enable cloud agents to interact seamlessly with external functions and environments.

In-Depth Analysis

Architectural Foundations: The Codex Harness as an Agent Engine

The introduction of the Agents API marks an architectural milestone in the evolution of autonomous AI systems. While previous generation interfaces focused largely on direct, stateless prompt-and-response interactions, the Agents API abstracts agent execution into a fully managed service. At the core of this system is OpenAI's Codex harness. By relying on the Codex harness, the Agents API provides the operational runtime and scaffolding needed to govern how agents execute commands, preserve context, and maintain operational stability during complex tasks. Instead of requiring engineering teams to construct bespoke scaffolding for execution loops, error handling, and runtime sandboxing, the platform supplies this infrastructure as a native, managed cloud offering.

Solving Session Persistence and Complex Orchestration

One of the most formidable hurdles in agent development has been managing long-running tasks. Traditional AI interactions degrade when stretched across prolonged timelines due to context window saturation, state drift, and infrastructure dropouts. The Agents API directly addresses this friction by providing managed support for orchestration and long-running sessions. In this environment, orchestration refers to the structured coordination of sub-tasks, planning phases, and response handling across time. Because the service manages the session lifecycle in the cloud, agents can run extended operations asynchronously, maintaining the continuity required to finish complex objectives without losing state.

Operationalizing Tool Use in the Cloud

Autonomous agents derive their utility from their ability to interact with outside software, databases, and APIs. The Agents API formalizes tool use as a native capability of the managed runtime. Through the Codex harness, the API oversees how models interpret tool specifications, execute calls, and process returned data. Centralizing tool use inside a managed cloud framework ensures that agents do not simply generate disconnected text, but instead execute programmatic actions safely and predictably. This integration establishes a robust execution loop where models observe, decide, call external tools, and continue their mission within a persistent operational boundary.

Industry Impact

The release of the Agents API reflects a broader structural evolution across the artificial intelligence sector: the migration from raw foundation models to managed autonomous runtimes. For developers and enterprises, building production-ready AI agents historically demanded substantial engineering overhead to maintain stateful servers, manage orchestration frameworks, and secure tool execution pipelines. By providing an end-to-end managed service powered by the Codex harness, OpenAI significantly lowers the barrier to deploying cloud-native agents.

Furthermore, this development solidifies "the agent harness" as a distinct layer in the modern software stack. Rather than building proprietary orchestration glue, development teams can rely on a managed standard to handle execution lifecycles, enabling them to focus their engineering resources on business logic, tool design, and domain-specific agent behaviors. As long-running cloud agents become standard components of automated business workflows, managed APIs of this nature will define how autonomous software interacts with digital infrastructure.

Frequently Asked Questions

What is the primary purpose of the Agents API?

The Agents API is a managed service developed by OpenAI that enables engineers to build, deploy, and launch cloud-based AI agents with minimal infrastructure management.

What role does the Codex harness play in the Agents API?

The Codex harness serves as the core underlying engine powering the Agents API, providing the critical orchestration framework required for session management, tool execution, and runtime stability.

What key features does the Agents API introduce for agent development?

The platform focuses on three primary operational capabilities: multi-step task orchestration, persistent support for long-running sessions, and robust tool use integration.

Read original article

Related News

Google Announces Gemini 4 Argon Frontier Model Restricting Initial Access to Trusted Cyber Defenders
Product Launch

Google Announces Gemini 4 Argon Frontier Model Restricting Initial Access to Trusted Cyber Defenders

Google has officially revealed Gemini 4 Argon, its latest frontier artificial intelligence model designed to deliver cutting-edge performance across complex enterprise workflows. Announced by Google DeepMind Senior Vice President and Chief AI Architect Koray Kavukcuoglu, the new system is built to excel in real-world software engineering, cybersecurity defense, and high-stakes enterprise knowledge tasks such as finance and legal operations. However, recognizing the unprecedented power and advanced capabilities of the system, Google is deliberately withholding a broad public release. Instead, the tech giant is restricting early access strictly to vetted, trusted cyber defenders. This cautious rollout strategy highlights the growing industry emphasis on defensive readiness and risk management as frontier AI systems reach higher levels of operational autonomy.

Product Launch

CrawlRaven Launches MCP Server on Product Hunt to Connect AI Agents Directly to SEO and Analytics Data

On September 30, 2026, developer Ayush Chaturvedi launched CrawlRaven MCP on Product Hunt, bringing a dedicated Model Context Protocol server to modern search engine optimization workflows. The new release bridges AI agents—including Claude, ChatGPT, and Cursor—directly with Google Search Console and Google Analytics 4 data through a secure, browser-based OAuth authentication flow. By deploying 13 read-only tools, CrawlRaven MCP eliminates repetitive spreadsheet exports, enabling AI assistants to natively surface ranked optimization opportunities, track slipping keyword queries, and analyze technical site audits through simple conversational prompts. The integration reflects the broader industry transition toward agent-driven data retrieval and automated marketing workflows, providing developers and SEO specialists with actionable search intelligence directly inside their daily developer environments.

OpenAI Launches GPT-6.1 Sol Nearing Astra Performance as Factual Errors Drop to 7.7 Percent
Product Launch

OpenAI Launches GPT-6.1 Sol Nearing Astra Performance as Factual Errors Drop to 7.7 Percent

OpenAI has officially launched GPT-6.1 Sol, a new artificial intelligence model that the company reports is nearing the performance capabilities of Astra. According to the reported data, the new model achieves notable improvements in accuracy, particularly when operating under low reasoning effort parameters. Specifically, benchmark measurements indicate that factual errors dropped significantly from 11.4% down to 7.7% in this operational tier. This measurable reduction in factual inaccuracies highlights OpenAI's continued technical focus on refining factual precision and reasoning reliability across different computational workloads. While comprehensive technical documentation and broader comparative metrics remain limited in the initial disclosure, the drop in error frequency represents a critical milestone for AI reliability in baseline reasoning workflows.