OpenAI’s New Agents API Could Change How SMEs Automate Work
OpenAI has released its Agents API in public beta, giving developers access to the managed harness and infrastructure used to run Codex. For a Malaysian SME, the important point is not the technical announcement alone. It is the possibility of deploying software agents that can handle multi-step work, use business tools, coordinate tasks and continue operating across longer sessions.
Instead of building every part of an agent system yourself, a developer can connect a model to tools, files, instructions and an execution environment through the API. OpenAI says the service is available to all developers in public beta, although data residency is currently US-only and Zero Data Retention is unsupported. (Source)
This matters if you are trying to automate customer enquiries, document processing, internal reporting, logistics coordination or software support with a small team. The technology is still in beta, so it should be approached as a controlled business experiment rather than an unattended replacement for your staff.
What Happened
OpenAI’s Agents API is a managed service built around the open-source Codex harness. The company describes the harness as the layer that manages context, tool use, subagents and reliable long-running execution. Developers can choose an OpenAI-hosted sandbox, run the agent in their own infrastructure, use a partner sandbox or operate without a sandbox. (Source)
The API is organised around four main concepts: an agent, an environment, a session, and events or items. An agent contains the model, instructions, tools and MCP servers available to it. An environment provides access to files, packages, skills and commands. A session is a durable agent instance that can receive tasks and provide updates. Events and items represent the information sent to the agent and the output it produces. (Source)
A typical session begins when you create it and provide a task. You can then monitor progress through streaming or webhooks, before sending a new task or steering the current one. OpenAI’s example creates an incident-investigation agent in one API call, connects it to an observability tool through MCP, enables up to three concurrent subagents and saves findings to a workspace. (Source)
The harness also includes automatic context compaction, tool search, programmatic tool calling and multi-agent support. In practical terms, this means an agent can retain the useful parts of a long task, load tool definitions when they are needed, process results in code and divide complex work among specialised subagents. (Source)
Why This Matters for Malaysian SMEs
Most SMEs do not need an agent to perform impressive demonstrations. You need fewer missed follow-ups, faster document handling and clearer visibility over daily operations. The Agents API could support workflows such as reading supplier documents, checking required fields, preparing a summary and sending an approval request to the right person.
Consider a Malaysian wholesaler receiving purchase orders through email, PDF and messaging applications. An agent could extract product codes, quantities and delivery dates, check the information against your inventory system and flag unusual requests for human review. It could prepare a draft response in English or Bahasa Malaysia, while leaving the final approval to your operations team.
A service company could use a similar workflow for customer support. The agent could review a support conversation, search an internal knowledge base, check a job record and prepare a suggested reply. If the issue requires a site visit or technical escalation, the agent could create a structured task instead of allowing the request to remain buried in a chat thread.
There are also useful applications for professional firms. An accounting, recruitment, compliance or consulting practice could ask an agent to organise client documents, identify missing information, compare versions and prepare a review checklist. The agent should not make final regulatory or financial decisions, but it can reduce repetitive preparation work so your experienced employees focus on judgement.
The practical opportunity is not to let an agent run your entire company. It is to give one carefully selected workflow a reliable assistant that can read, check, coordinate and prepare work for approval.
Possible SME Use Cases
| Business area | Potential agent task | Human checkpoint |
|---|---|---|
| Sales | Qualify enquiries and prepare follow-up drafts | Salesperson approves message and customer status |
| Operations | Compare orders, stock records and delivery information | Operations lead approves exceptions |
| Customer service | Search knowledge sources and suggest replies | Staff handles sensitive or unusual cases |
| Administration | Extract fields from forms and organise files | Authorised employee confirms key details |
| IT support | Investigate alerts and produce incident summaries | Technical owner approves changes or recovery actions |
What You Need to Check Before Using It
The largest issue for many Malaysian businesses is data handling. OpenAI’s announcement states that Agents API data stays US-only and that Zero Data Retention is not supported. (Source) If your workflow contains personal information, confidential contracts, health information, customer identity details or regulated records, you should obtain professional advice and review your contractual obligations before sending the data to the service.
Start with low-risk information. For example, use anonymised support tickets, fictional purchase orders or internal documents that do not contain sensitive personal details. Confirm what the agent can access, what it can change and which actions require approval. Avoid giving an early pilot permission to delete records, send binding commitments or modify financial data.
You should also map the workflow before asking a developer to build it. Write down the trigger, the information required, the systems involved, the expected result, the failure conditions and the person responsible for approval. This will reveal whether an agent is appropriate or whether a simpler rule-based automation would be easier to operate.
The Bigger Picture
The Agents API sits between a fully managed agent service and a developer-controlled software system. OpenAI’s comparison describes the Agents API as having lower integration effort, saved session configuration and a managed Codex harness. The Agents SDK runs inside your application, while the Responses API gives your application more direct control over history, orchestration and execution. (Source)
That choice is important for an SME. If you want to test one internal workflow quickly, a managed service may reduce the engineering work. If your process requires strict control over infrastructure, data location, audit trails or custom orchestration, your developer may recommend a self-hosted or more controlled approach instead.
OpenAI has published customer-reported results from Ciridae, SafetyKit, Hypha and Nash.ai, including changes in evaluation scores, response failures, workflow performance and long-running agent usage. These figures are vendor-supplied rather than independent benchmarks, so you should not assume the same outcome for your business. (Source)
Your next step should be modest: select one repetitive process, define a measurable target, remove sensitive data from the first test and keep a person in the approval loop. Track accuracy, handling time, escalations and mistakes. If the pilot performs reliably, expand it gradually and document the controls.
For Malaysian SMEs, the Agents API is a signal that business automation is moving from simple question-and-answer tools towards systems that can coordinate several steps. The winners will not be the businesses that deploy the most agents. They will be the businesses that choose sensible workflows, protect customer information and make automation accountable from the beginning.
Ready to Streamline Your Operations?
Technology moves fast. Your operations should keep up. AutoRunBiz builds AI systems that run your daily workflows — from WhatsApp order capture to accounting. Book a free 15-min ops audit →