Join the event that trusts business leaders for almost two decades. VB Transform brings together people who build a real business AI strategy. Learn more
HAS Venturebeat conference transform 2025Olivier Godment, responsible for the product of the OPENAI API platform, provided behind the scenes of how business teams adopt and deploy large-scale AI agents.
In a 20 -minute round table, I hosted exclusively with God, the former Stripe and current researcher Boss Api Openai unpacked the latest OpenAi developer tools – API responses and SDK agents – while highlighting the real world models, safety considerations and examples of cost return from early adopters like Stripe and Box.
For business leaders unable to attend the live session, here are the 8 best dishes to remember:
Agents quickly go from prototype to production
According to God, 2025 marks a real change in the way AI is deployed on a large scale. With more than a million monthly active developers now using the OPNAI API platform in the world and the use of tokens up from 700% from one year to the next, AI goes beyond experimentation.
“It’s been five years since we launched GPT-3 mainly … and guy, the last five years have been quite wild.”
Divination stressed that current demand no longer concerns chatbots. “The use cases move from simple questions and answers to really use cases where the application, the agent, can do things for you.”
This change prompted Openai to launch two main tools oriented towards developers in March: the API APPLIES and the Agent SDK.
When to use unique agents vs undergarget architectures
A major theme was the choice of architecture. Divination noted that the unique agent loops, which encapsulate access and the full context of tools in a model, are conceptually elegant but often unsatisfied on a large scale.
“Building specific and reliable unique agents is difficult. As, it’s really difficult.”
As complexity increases – more tools, more possible user inputs, more logic – teams often move to modular architectures with specialized sub -agents.
“A practice that has emerged is to essentially break down agents into several sub-agents … You would separate concerns as in software.”
These sub-agents operate as roles in a small team: a sorting agent classifies the intention, level 1 agents manage routine problems and others increase or solve on-board cases.
Why the API of responses is a change of steps
Divination has positioned the API of responses as a fundamental evolution of developers’ tools. Previously, the developers have manually orchestrated model call sequences. Now, this orchestration is managed internally.
“The API of responses is probably the largest new layer of abstraction that we have since introduced GPT-3.”
It allows developers to express the intention, not just to configure model flows. “You care to return a very good answer to the customer … The response API mainly manages this loop.”
It also includes built -in capacities for knowledge recovery, web research and functions of functions – companies that companies need works of real agents.
Observability and safety are integrated
Safety and compliance were in the lead. Godment has cited key railings that make OPENAI viable stack for regulated sectors such as finance and health care:
- Policy refusal
- SOC-2 journalization
- Data residence support
The evaluation is the place where divination sees the greatest gap between demo and production.
“My hot intake is that the assessment of the model is probably the largest bottleneck for the massive adoption of AI.”
OPENAI now includes tracing and evaluation tools with the API battery to help teams define what success is like and follow how agents work over time.
“Unless you invest in evaluation … It is really difficult to strengthen this confidence, this confidence that the model is correct, reliable.”
The early king is visible in specific functions
Some cases of use of the company already provide measurable gains. Shared example of God of:
- Bandwhich uses agents to speed up the handling of invoices, the reports “Invoice resolution 35% faster”
- Boxwho launched knowledge assistants that allow “Sorting of touch tickets”
Other high -value use cases include customer support (including voice), internal governance and knowledge assistants to navigate dense documentation.
What to be launched in production
Divination highlighted the human factor in successful deployments.
“There is a small fraction of very high -end people who, each time they see a problem and see a technology, they run there.”
These internal champions do not always come from engineering. What unites them is persistence.
“Their first reaction is, ok, how can I operate it?”
OPENAI sees many initial deployments motivated by this group – people who have pushed the first use of chatgpt in the company and are now experimenting with full agent systems.
He also underlined a gap that many ignore: the expertise of the domain. “Knowledge in a business … does not live with engineers. It lies in the OPS teams. “
Making the construction tools accessible to non-development is an OPENAI challenge aims to take up.
What is the next step for corporate agents
Divination offered an overview of the roadmap. OPENAI actively works:
- Multimodal agents who can interact via text, voice, images and structured data
- Long -term memory To maintain knowledge through the sessions
- Cross orchestration To support complex and distributed IT environments
These are not radical changes, but iterative layers that develop what is already possible. “Once we have models that may think not only for a few seconds, but for minutes, for hours … this will allow fairly breathtaking cases of use.”
Final word: reasoning models are subcontracted
Divination closed the session by reaffirming its conviction that the models compatible for reasoning – those which can reflect before responding – will be the real catalysts of long -term transformation.
“I always have a conviction that we are roughly at the GPT-2 or GPT-3 maturity of these models … We always explore the surface of what the reasoning models can do.”
For corporate decision -makers, the message is clear: the infrastructure for agent automation is here. What matters now is to build a targeted use case, empower interfunctional teams and be ready to iterate. The next phase of value creation does not reside in new demos, but in sustainable systems, shaped by the needs of the real world and operational discipline to make them reliable.
