But Venturebeat Transform 2025 ConferenceOlivier Godem, head of the product for the API platform of OpenAII, saw a rear scenes how enterprise teams are adopting AI agents and deploying AI agents on a scale.

In a 20-minute panel discussion, I hosted with Godem, especially, former strip researcher and current OpenAI API boss unpacked OpenEII’s latest developer tools-reactions API and agent SDK-while the actual-world pattern, safety views and early adopted were highlighted.

For enterprise leaders unable to participate in the session Live, here are the top 8 most important takeaways:

Agents are fast moving from prototype to production

According to Godem, there is a real change in 2025 how AI is being deployed on a scale. With more than one million monthly active developers, the API platform of OpenAII is now used globally, and the use of tokens is 700% year -to -year, AI is moving beyond the use.

“It has been five years because we have essentially launched the GPT-3 … and the man, the last five years have been very wild.”

Godem insisted that the current demand is not yet about chatbots. “AI use cases are actually moving to use cases from simple Q&A where apps, agents, can do goods for you.”

This innings inspired Openai to launch Two major developer-focus equipment in March, Response API And this Agent SDK,

Use when single agents versus sub-agent architecture

A major topic was liked by architectural. Godment stated that single-agent loops, which encounter complete equipment access and reference in a model, are ideologically elegant, but often impractical on a scale.

“The manufacture of accurate and reliable single agents is difficult. As, it is really difficult.”

As the complexity increases-more equipment, more potential user inputs, more arguments-teams often move towards modular architecture with special sub-agents.

“An exercise that has emerged, essentially to break agents into several sub-agents … you will separate concerns like software.”

These sub-agents act like roles in a small team: a triangle agent classifies intentions, tier-one agents handle regular issues, and solve other edge or edge matters.

Why reactions API is a step change

Godem deployed APIs in the form of a fundamental development in developer tooling. Previously, developers orchestrained the sequences of model calls manually. Now, that orchestration is handled internal.

“Reactions API is probably the biggest new layer of abstraction offered since much more GPT -3.”

This allows developers to express intentions, not just the model flow configures. “You care about giving the customer a really good response … The response API essentially handles that loop.”

It also includes the inherent capabilities for knowledge recover, web discovery, and function calling-which enterprises are required for real-world agent workflow.

Created in observation and security

Security and compliance were the top of the brain. Godement cited major railings, which make Openai stacks viable for regulated sectors such as finance and healthcare:

• Policy-based refinance
• SoC-2 logging
• Data residence support

Evaluation is the place where Godem sees the biggest difference between demo and production.

“My hot tech is that model assessment is probably the biggest hurdle to adopt AI on a large scale.”

Openai now includes tracing and eval tools with API stack to help teams to define what success looks and tracks how agents perform over time.

“Until you invest in evaluation … it is really difficult to build that belief, it is believed that the model is accurate, reliable.”

Initial ROI appears in specific functions

Some enterprise use cases are already giving an average benefit. Shared example from Godem:

• stripWhich uses agents to deal with challans, speed up reporting “35% fast invoices resolution”
• boxWho launched knowledge assistants that enables “Zero-Tach Ticket Triats”

Other high-value use cases include customer aid (including voice), internal governance and knowledge assistant for navigating dense documents.

What does it take to launch in production

Godem emphasized the human factor in successful deployment.

“There is a small fraction of very high-end people, which whenever they see a problem and see a technique, they walk on it.”

These internal champions do not always come from engineering. The one who units them is firmness.

“His first response is, well, how can I work it?”

Openai sees several initial deployment run by this group – those who have used early chats in the enterprise and are now experimenting with full agent systems.

He also told many neglect: Domain expertise. “Knowledge in an enterprise … does not lie with engineers. It is with OPS teams.”

It is a challenge to make the agent-beding tool accessible to non-decayers, which is to address OpenAII.

What’s next for enterprise agents

Godem gave a glimpse in the roadmap. Openai is actively working:

• Multiple agent It can interact through text, voice, drawings and structured data
• Long term memory To maintain knowledge in sessions
• Cross cloud orchestration To support the complex, distributed IT environment

These are not radical changes, but recurrence layers that are already possible. “Once we have models that can think not only for a few seconds but for minutes, for hours … it’s going to enable some beautiful mind-making cases.”

Last Words: Logic models are underhipd

Godment confirmed its confidence to close the session that logic-competent models-which could reflect before answering-would be the correct ambassadors of degradable changes.

“I still believe that we are very high at the GPT-2 or GPT-3 level of maturity of those models … we are still scratching the surface what logic models can do.”

For enterprise decision makers, the message is clear: the infrastructure for agent automation is here. Now what matters is creating a concentrated use case, strengthening the cross-functional teams, and is ready for recurring. The next stage of value construction is not in the novel Demo-but in sustainable systems, the real world needs and the operation discipline shaped to make them reliable.