發生了什麼事
BigGo Finance reports that OpenAI launched the Agents API in public beta and is making the execution framework used with Codex available to external developers. The framework manages context, tool use, task sequencing, multi-agent workflows and sandboxed code execution. The report says the API has no separate usage fee, but model, tool and sandbox charges still apply.
According to BigGo Finance, OpenAI announced the Agents API in public beta on the 10th and made the Codex “harness” available to developers. The report describes the harness as an execution framework that manages task context, invokes tools and coordinates sequences of work rather than simply calling a model.
The report says long-running tasks can automatically compress earlier context, while multiple sub-agents can work in parallel and a primary agent can combine their results. It also says developers can use web search, MCP connections and custom tools, run operations concurrently, and choose OpenAI-managed sandboxes, their own infrastructure or supported environments from Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel.
BigGo Finance says the Agents API has no separate usage fee, but charges accrue for underlying model tokens, tools and OpenAI-provided sandboxes. It reports customer results from SafetyKit, Hypha, Cirridae and Nash, including lower costs, fewer failures, improved evaluation scores, reduced latency and large-scale long-running deployments. These figures are attributed to the report and are not independently confirmed here.
為什麼這很重要
The reported change could lower the engineering cost of building agents that do more than generate responses. By providing infrastructure for persistent context, tool coordination, parallel sub-agents and code execution, OpenAI is competing at the workflow layer as well as the model layer. That matters for companies deciding whether to build agent runtimes themselves or depend on a managed platform. The source does not independently verify the reported customer outcomes, and it does not establish general availability beyond the public beta or show how the system performs across broader workloads.
If the reported capabilities work as described, developers may be able to spend less effort assembling the basic runtime needed for agents that read and write files, execute code and continue work over extended periods. This shifts an important part of competition from model quality alone toward execution reliability and orchestration.
The approach could also create platform dependence: OpenAI would maintain the harness as its models change, while customers could connect external sandboxes and tools. That may simplify maintenance but makes the API’s security controls, data handling, portability and total operating cost important purchasing questions.
The source presents early customer metrics but supplies no methodology, baseline details or independent validation. Those results should therefore be treated as reported case studies rather than evidence of performance across the wider developer market.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
接下來看什麼
Developers will need to assess reliability, cost, security and portability in real deployments. The report says data residency is currently limited to the United States and that is unavailable, which may restrict use with sensitive information. The source does not provide a general availability date, a complete pricing schedule, model eligibility details or independent testing of the reported performance gains.
The report says United States data residency is the current limit and is not offered. Organizations handling confidential data will need to determine whether those conditions meet their legal, contractual and internal security requirements.
The source does not document a general availability date, complete pricing, rate limits, supported models or the boundaries of the public beta. It also does not independently confirm the cited customer deployments or performance improvements.
Future updates may show whether developers adopt OpenAI’s managed execution layer, continue building their own runtimes, or use competing infrastructure providers. Reliability on long-running tasks and the cost of repeated model and tool calls will be particularly consequential.