OpenAI’s Agents API Brings the Codex Harness to Developers

OpenAI has put its Agents API into public beta, making the orchestration harness behind Codex available as a managed developer service. The pitch is straightforward: developers configure the work an agent should do, the model and tools it can use, and where it should run; OpenAI operates the harness that coordinates the work. 15

What OpenAI launched

The Agents API provides a way to build and run cloud agents on OpenAI’s managed Codex harness. Rather than requiring application teams to operate every layer of an agent runtime themselves, the service handles model calls, context management, tool use, and coordination among agents. 15

Developers can select the compute environment for their agents:

• an OpenAI-managed sandbox;
• their own infrastructure; or
• sandbox partners including Cloudflare, DigitalOcean, Oracle, and Vercel. 15

That separation matters for teams that want managed orchestration but retain a choice over where code and workloads execute.

What the managed harness does

Agent applications often need more than a model endpoint: they must preserve useful context over long sessions, decide which tool definitions to load, execute dependent or parallel tool calls, and delegate parts of a task to specialized agents. OpenAI says the API includes capabilities aimed at those operational problems. 15

They include:

• Automatic context compaction for longer-running sessions;
• Tool search, which can load relevant tool definitions when needed;
• Programmatic tool calling for chained and parallel operations; and
• Concurrent subagent delegation, with separate contexts for subagents. 15

The practical implication is that teams can concentrate more on task design, tool access, guardrails, and product behavior, rather than maintaining the underlying orchestration layer. It also means accepting greater dependence on OpenAI’s managed runtime and its evolving product roadmap.

Pricing, availability, and the beta caveat

OpenAI says the Agents API is available to all developers in public beta, with no separate Agents API charge beyond standard model-token and tool costs. The company says the harness is based on the open-source Codex codebase, while OpenAI hosts and maintains the managed service. 15

Public beta availability should not be confused with a final production contract. OpenAI says it will refine the product using developer feedback before general availability, so teams evaluating it should validate operational requirements, portability, and performance for their own workloads. 15

OpenAI highlighted one early customer result from Ciridae: the company reported an evaluation score increase from 0.71 to 0.85 and a fourfold latency reduction for subagent workflows. Those are customer-reported results, not an independently published benchmark. 15

A managed alternative to building an agent runtime

The launch places OpenAI more directly in the market for agent frameworks and orchestration platforms. Its offering is not just model access: it combines a Codex-derived harness, execution-environment options, context management, tool optimization, and multi-agent coordination in one managed layer. 15

For developers, the trade-off is clear. Building on the API may reduce infrastructure work and accelerate deployment, while a self-managed stack can offer more direct control over implementation and operational changes. The right choice depends on how much control a team needs over its runtime versus how much maintenance it wants to avoid.

Two related launches: enterprise data and voice agents

OpenAI also introduced a Data agent in ChatGPT Work. It is designed to connect to approved company data, investigate questions, create interactive dashboards, and support follow-up actions through conversation rather than requiring users to manually consolidate data or write queries.

According to OpenAI, the Data agent supports sources such as Amazon Redshift, Google BigQuery, Databricks, MongoDB, Snowflake, Google Drive, and SharePoint, along with business-intelligence tools including Power BI, Tableau, Sigma, ThoughtSpot, Omni, and Oracle BI. Results can be shared through Slack or email, and the service is intended to respect existing table-, row-, and column-level permissions.

The company also released GPT-Live-1 in the API, a full-duplex voice model that can listen and speak at the same time and handle interruptions. OpenAI says it can hand off deeper reasoning and actions to paired text models and tools. 14

Yelp has integrated GPT-Live-1 into Yelp Host and Hatch for restaurant calls. OpenAI says those products have handled more than one million calls since October 2025; that figure is the company’s reported usage claim. 14

The takeaway

The Agents API turns the coordination system OpenAI uses for Codex into a public, managed building block. Developers still need to define the agent’s task, permissions, tools, and evaluation approach, but OpenAI is taking responsibility for much of the difficult runtime plumbing. 15

For teams that want to ship agent workflows without operating their own orchestration harness, the beta offers a lower-maintenance path. For teams with strict infrastructure, portability, or control requirements, the key question is whether that convenience outweighs the dependency on a managed platform that is still being refined.