OpenAI Releases Agents API and Data Agent: The Infrastructure Layer for Production AI Agents

On 10 September 2026, OpenAI made two moves that change how enterprises will build and deploy AI agents. The company launched its Agents API in public beta, bringing the orchestration harness behind Codex to developers through a managed API. On the same day, it introduced a Data Agent inside ChatGPT Work that lets organisations analyse connected company data and documents through natural language.

Taken together, the two releases show OpenAI shifting from "here is a powerful model" to "here is the infrastructure to run agents in production." The Agents API handles context management, tool calling, subagent delegation and long-running sessions. The Data Agent gives non-technical users a way to query structured and unstructured enterprise data without writing SQL. For buyers and suppliers in the agentic AI market, the signal is clear: the platform layer is maturing, and the competitive battleground is moving from raw model capability to operational reliability.

What the Agents API does

The Agents API is not a model endpoint. It is a managed harness that hosts, orchestrates and governs long-running AI agents. OpenAI has operated this infrastructure internally for Codex and ChatGPT Work. Now it is exposing it to external developers.

At its core, the API provides four capabilities. First, environment choice: developers can run agents in an OpenAI-hosted sandbox, on their own infrastructure, or through partner environments including Vercel, DigitalOcean, Cloudflare, E2B, Modal and others. This is important for enterprises with data residency or compliance requirements that cannot send everything to a third-party cloud. Second, context management: the API automatically compacts earlier context as sessions approach token limits, preserving the information the agent needs without manual intervention. This matters for agents that run for hours or days, where context windows would otherwise overflow and the agent would lose track of its mission. Third, tool orchestration: the API supports tool search to load relevant definitions on demand, parallel tool calling for efficiency, and programmatic filtering so agents only bring relevant results back into context. It works with MCP servers, custom functions and built-in tools. Fourth, multi-agent support: complex tasks can be broken into pieces and delegated to subagents that work in parallel, each maintaining independent context while a parent agent coordinates results.

OpenAI is also open-sourcing the core Codex harness that powers the API, giving developers visibility into the orchestration logic. The company says there are no additional fees for the Agents API beyond the tokens and tools the agent consumes.

What the Data Agent does

While the Agents API is developer-facing, the Data Agent is aimed at business users. Embedded inside ChatGPT Work, it connects to company data sources and lets users ask questions in natural language. A finance analyst can ask for quarterly variance analysis. A sales manager can request pipeline coverage by region. An operations lead can query supply chain metrics across multiple systems.

The significance is not that the agent can generate a chart. It is that it understands the data model, respects access controls, and operates within the same ChatGPT Work environment where employees already work. That removes the integration friction that kills most enterprise analytics projects: users do not need a new login, a new interface, or a new workflow. They ask a question and get an answer, complete with sourcing and the ability to drill down.

OpenAI has also published data showing how enterprise usage is shifting. Since February 2026, weekly active Codex users grew 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing, compared with 5-fold in engineering. Enterprise AI adoption is no longer a developer story. It is a general knowledge-work story. The Data Agent is designed for that audience.

Why the timing matters

The dual launch lands at a pivotal moment. Eight days earlier, Broadcom had unveiled AgentMinder, a runtime governance layer for enterprise AI agents. In July, Cisco announced it would deploy AI agents to all 90,000 employees. PwC launched an agentic front-office practice with OpenAI. The AvePoint State of AI 2026 report found that 88% of organisations suffered an AI agent-related security incident in the past year, and 86% delayed deployments citing governance concerns.

The pattern is consistent: enterprises want agents, but they need infrastructure that handles orchestration, context, tools, governance and data access in a single operational layer. OpenAI's response is to package the infrastructure it built for its own products and offer it as a platform. The question for buyers is whether they want to rent that infrastructure from OpenAI, or build or buy alternatives that give them more control over cost, data flow and vendor lock-in.

What it means for enterprise buyers

The Agents API and Data Agent carry three direct implications for organisations evaluating agentic platforms.

Agent orchestration is becoming a commodity. The core challenges of context management, tool calling and subagent coordination are hard engineering problems that most enterprises should not build from scratch. OpenAI is now offering a managed solution for those problems, which means buyers can shift engineering effort from infrastructure to application logic. The cost is dependency on OpenAI's pricing, availability and roadmap. Buyers should evaluate whether the time saved on orchestration outweighs the strategic risk of platform concentration.

Data access controls will become the new bottleneck. The Data Agent is only as good as the data it can reach and the permissions it respects. Enterprises that have clean data models, consistent access policies and well-documented APIs will see immediate value. Those with fragmented data estates, shadow databases and inconsistent governance will find that the agent surfaces problems that already existed but were hidden behind manual reporting workflows. The Data Agent does not fix dirty data. It exposes it.

Hybrid deployment models are now realistic. The Agents API's support for self-hosted and partner-hosted sandboxes means enterprises can keep sensitive workloads on premises or inside their own VPC while still benefiting from OpenAI's orchestration layer. This is a meaningful shift from the early days of API-only AI, where every prompt left the building. For regulated industries and government contractors, hybrid deployment may be the difference between running a pilot and getting production sign-off.

What it means for suppliers

For companies building agentic products or services, the OpenAI releases create both opportunity and competitive pressure.

The barrier to building agents is lower. Startups and mid-market software vendors can now use the Agents API to handle orchestration and focus their engineering on domain-specific tools, workflows and user interfaces. This accelerates time to market and reduces the capital required to launch an agentic product. The flip side is that differentiation becomes harder when everyone is using the same orchestration layer. Suppliers need to build defensible value in data, domain expertise or user experience rather than infrastructure.

Enterprise sales cycles will increasingly include infrastructure evaluation. Buyers will ask whether a supplier's agent runs on the Agents API, a competing orchestration layer, or custom infrastructure. Each choice carries implications for cost predictability, data residency, latency and vendor risk. Suppliers should be prepared to explain their architecture choice and the trade-offs it entails. Those that can articulate a clear rationale will win trust faster than those that treat infrastructure as an implementation detail.

The partner ecosystem is expanding. OpenAI has integrated with Cloudflare, Vercel, DigitalOcean, Modal, E2B, Daytona, Blaxel, Oracle and Runloop for sandbox hosting. This creates opportunities for infrastructure partners to differentiate on cost, performance or compliance capabilities. It also means suppliers have more choice about where and how their agents run. The ecosystem effect is real: the more deployment options, the more comfortable enterprises become with production deployment.

The broader context

OpenAI's moves are part of a broader shift in the AI industry from model competition to platform competition. The gap between frontier models is narrowing. Anthropic's Claude, Google's Gemini and open-source alternatives like Nous Research's Hermes are all capable of powering sophisticated agents. What distinguishes winners in the enterprise market is no longer raw benchmark performance but operational reliability, integration depth, governance support and total cost of ownership.

The Agents API positions OpenAI as an infrastructure provider, not just a model provider. That is a higher-margin, stickier business, but it also brings higher expectations for uptime, security certifications and enterprise support. OpenAI's public beta status signals that the product is still stabilising. Enterprises should treat early deployments as bounded pilots, not enterprise-wide rollouts, until the service reaches general availability and publishes SOC 2 or similar compliance documentation.

The Data Agent, meanwhile, competes with a crowded field of enterprise analytics and business intelligence tools. Its advantage is integration with ChatGPT Work, which many enterprises already use. Its disadvantage is that analytics accuracy depends on data quality, schema stability and access control rigour, none of which OpenAI can solve for a customer. The product will shine in organisations that have already invested in modern data infrastructure. It will frustrate those that have not.

The Agentic Expo angle

Agentic Expo exists at the intersection of capability and trust. OpenAI's dual launch advances capability significantly. The trust layer, governance, data control and operational verification, still sits with the buyer and their architecture choices.

By March 2027, the market will have moved past the question of whether agents can work, and into the question of which infrastructure stack to standardise on. The exhibitors and speakers at Agentic Expo will represent every layer of that stack: the model providers, the orchestration platforms, the governance vendors, the data infrastructure companies and the consultancies stitching them together. OpenAI has just advanced the state of the orchestration layer. Our event will show what buyers and suppliers are doing with it.