OpenAI turned the agent infrastructure behind Codex into a single managed API
On September 10, 2026 OpenAI released the Agents API as an open beta for all developers. According to the official developer documentation, the API gives applications access to the Codex harness through an OpenAI-managed interface; sessions, subagent orchestration, context compaction and recovery sit with OpenAI. As Investing.com describes it, the task, model, tools and compute environment stay with the developer. The API itself carries no extra fee, and data residency is US-only for now.
The Codex harness now sits behind an API that OpenAI operates
OpenAI's developer documentation defines the Agents API as giving an application access to the Codex harness through an OpenAI-managed API. MarkTechPost and Investing.com, both dated September 10, 2026, report that the API opened to all developers as a beta that day. The beta status shows in the documentation too: requests carry an OpenAI-Beta header set to agents=v1, the SDK examples use the beta.agents namespace, and the sample model is gpt-6-astra. As Investing.com describes the division of labor, the developer defines a task, a model, tools and a compute environment, while OpenAI operates the Codex harness and the agent infrastructure underneath it. Per the official documentation, the harness manages sessions, orchestration, context compaction and recovery; supports programmatic tool calling, MCP servers and web search; can break work into subtasks and delegate them to subagents; and can resume a session where it left off. The architecture document adds that the harness can call remote MCP tools directly and that OpenAI keeps the session for later work, with progress delivered by streaming or webhook.
Five features, three environment types, nine sandbox partners and no API surcharge
MarkTechPost lists five core features: context compaction (earlier context is compressed automatically as a session nears its limit), tool search (tool definitions load only when needed, saving tokens), programmatic tool calling (parallel execution and chaining), subagents (delegation with an independent context) and persistent sessions (configuration, turns and items are stored). The official documentation defines the compute environment through three environment.type values: none, for question answering and external service access without a sandbox; openai_hosted, for running scripts and editing files; and self_hosted, for custom infrastructure and software requirements. The built-in bash and apply-patch tools exist only in the OpenAI-hosted environment. According to MarkTechPost, teams bringing their own infrastructure connect through codex exec-server over WebSocket, and nine partners provide sandboxes: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. Investing.com adds that the partners offer managed environments, virtual private cloud deployments and varied CPU, GPU and memory configurations, and that an open-source version of the harness is available in OpenAI's public codebase. Pricing has three lines in the official documentation: model usage at the selected model's API rates, OpenAI tools at standard rates, OpenAI-hosted sandboxes at standard container rates. According to Investing.com, no additional fee applies for the API itself.
Agent infrastructure is turning from something you build into something you rent
The hardest part of keeping an agent in production is not the model but the infrastructure around it. Session limits, subtask distribution, recovery and sandboxing are each engineering work of their own; the Agents API turns that layer into a service OpenAI operates. Early customer results are vendor claims: according to Investing.com, Ciridae CTO Jack Weissenberger reported an evaluation score rising from 0.71 to 0.85 and a fourfold drop in latency in subagent flows; MarkTechPost relays a 60 percent drop in cost per case at SafetyKit and an 86 percent reduction in failed responses at Hypha. The limits are just as clear: per the official documentation, the Agents API currently supports data residency only in the United States, does not support Zero Data Retention (ZDR), and a self-hosted sandbox does not make it ZDR-eligible; MarkTechPost notes this limits regulated workloads for now. Investing.com reports that OpenAI plans to refine the API with feedback before general availability; the sources give no GA date or timeline for residency outside the United States.
For businesses in Türkiye: a clear opportunity, with a KVKK check first
The boundary first: none of the sources mentions Türkiye or the region; the only geographic restriction is US-only data residency. The practical meaning is clear. For agencies and SMEs in Türkiye, the Agents API means running a production agent without building the sandbox, session management and subagent orchestration layers themselves; the pricing model has no separate license, only token, tool and container usage. Having Cloudflare, DigitalOcean and Vercel, providers many teams in Türkiye already know, among the nine sandbox partners makes adoption easier. The part that needs care is data: US-only residency and the absence of ZDR mean a business processing personal data under KVKK must run a cross-border transfer assessment, and a self-hosted sandbox does not change that, since the official documentation says it does not make the API ZDR-eligible. A sensible order: start with workloads that contain no personal data (internal reporting, code generation, public content processing); then run the transfer assessment for scenarios that need personal data, and make deleting unneeded sessions and artifacts a routine.
The UNALSOFT view
In our Agentic AI solutions work we look less at the harness than at what surrounds it: the agent's task, its tools, when it hands over to a person and where the data lives. The Agents API takes the infrastructure part of those questions to OpenAI; the decisions stay with the business. So our recommended order does not change: write down the task and tool set first, then classify which data enters the agent, and choose infrastructure last. Where a platform stores data changes over time; a well-defined task moves to any infrastructure.
Thinking about putting an agent into production in your business?
Let's define the task, the tool set and the data boundaries together and see whether a managed infrastructure like the Agents API fits. A short conversation is enough to start.