Agent toolkits become standard issue. What your software vendor will do with that

OpenAI released the Responses API today, 11 March 2025, together with an open-source Agents SDK: tooling for building software that searches, reads files, operates a browser and coordinates several steps without a person at each one. Notably, the SDK can orchestrate models from other providers too.

This is developer infrastructure. No SME will touch it directly. It matters because it changes what your existing suppliers can build, and how quickly.

What actually changes for you

Until now, an agent feature in a business application was a serious engineering project. The vendor had to build the orchestration, the retries, the tool handling, the logging — before writing a line of anything specific to your industry.

With standard toolkits, that plumbing is a dependency rather than a project. Expect a visible increase in agent features arriving in the software you already pay for, from vendors who could not have built them last year.

Some of those features will be excellent. Some will be a weekend’s work shipped with confidence.

The three questions to ask a vendor this year

• What can it do without asking me? The only question that matters. Draft, sort, summarise and flag are one category. Send, pay, delete and commit are another. Make the vendor state which side their feature sits on, in writing.
• Where does it get stuck? A considered agent has an explicit stopping condition: it gives up, tells someone and leaves a record. One that retries silently until it succeeds is a system that will eventually succeed at the wrong thing.
• What is in the log? If the agent takes five steps and step three was wrong, can you see step three? If the answer is no, you cannot debug it and neither can they.

The detail worth noticing

The toolkit orchestrates models from several providers, not only its author’s. That is the more interesting signal: the competitive layer is moving up from the model to what is built on top of it.

For a buyer that is good news. It makes « can we change the model underneath? » a question with a plausible yes, where a year ago it had none.