OpenAI API Lead Details Enterprise Success with Agents SDK and Responses API

At VentureBeat's Transform 2025 conference, Olivier Godement, Head of Product for OpenAI's API platform, offered an inside perspective on how enterprise teams are implementing and scaling AI agents.

In an exclusive 20-minute panel I hosted with Godement, the former Stripe researcher and current OpenAI API lead detailed the company's latest developer tools—the Responses API and Agents SDK—while showcasing real-world applications, security measures, and return-on-investment examples from early adopters such as Stripe and Box.

For enterprise leaders who missed the live session, here are the eight most important insights:

Agents Are Rapidly Moving From Prototype to Production

Godement highlighted that 2025 represents a significant shift in the large-scale deployment of AI. With over one million monthly active developers globally using OpenAI's API platform and a 700% year-over-year increase in token usage, AI is advancing beyond the experimental stage.

"It's been five years since we first launched GPT-3… and the journey since then has been truly remarkable."

He pointed out that current demand extends well beyond chatbots. "AI applications are evolving from simple Q&A to use cases where the agent can actively perform tasks on your behalf."

This transition led OpenAI to introduce two key developer tools in March: the Responses API and the Agents SDK.

When to Use Single Agents vs. Sub-Agent Architectures

A central topic was architectural decision-making. Godement explained that while single-agent loops—which combine full tool access and context within one model—are elegant in theory, they often prove challenging to implement at scale.

"Creating accurate and dependable single agents is difficult. It's genuinely tough."

As systems grow more complex—with additional tools, diverse user inputs, and intricate logic—teams frequently adopt modular designs using specialized sub-agents.

"A common approach is to break down agents into multiple sub-agents… applying software engineering principles like separation of concerns."

These sub-agents operate like roles within a team: a triage agent categorizes requests, tier-one agents manage routine tasks, and others address complex or unusual cases.

Why the Responses API Is a Step Change

Godement described the Responses API as a fundamental advancement in developer tools. Previously, developers had to manually manage sequences of model calls. Now, this coordination is handled automatically.

"The Responses API is arguably the most significant layer of abstraction we've introduced since GPT-3."

It enables developers to focus on intent rather than configuring model workflows. "Your goal is to deliver an excellent response to the customer… the Response API manages that entire process."

The API also includes built-in features for knowledge retrieval, web search, and function calling—essential components for real-world enterprise agent workflows.

Observability and Security Are Built In

Security and compliance were emphasized as critical considerations. Godement highlighted key safeguards that make OpenAI's platform suitable for regulated industries like finance and healthcare:

• Policy-based refusals
• SOC-2 logging
• Data residency support

He identified evaluation as the primary challenge in moving from demonstration to production.

"In my view, model evaluation is likely the biggest obstacle to widespread AI adoption."

OpenAI now integrates tracing and evaluation tools within the API stack, helping teams define success metrics and monitor agent performance over time.

"Without investing in evaluation… it's difficult to build the trust and confidence needed to ensure model accuracy and reliability."

Early ROI Is Visible in Specific Functions

Certain enterprise applications are already showing measurable benefits. Godement provided examples from:

• Stripe, which uses agents to speed up invoice processing, reporting "35% faster invoice resolution"
• Box, which deployed knowledge assistants enabling "zero-touch ticket triage"

Other high-impact use cases include customer support (including voice-based systems), internal governance, and knowledge assistants for navigating complex documentation.

What It Takes to Launch in Production

Godement stressed the importance of human initiative in successful deployments.

"A small group of highly motivated individuals see a problem and a technological solution, and they dive right in."

These internal advocates don't always come from technical backgrounds. What they share is determination.

"Their immediate response is, 'How can I make this work?'"

OpenAI observes that many initial deployments are driven by this cohort—early adopters of ChatGPT in corporate settings who are now exploring full agent systems.

He also highlighted an often-overlooked factor: domain expertise. "The essential knowledge within an enterprise… resides not with engineers, but with operational teams."

Enabling non-developers to use agent-building tools is a key challenge OpenAI is working to solve.

What’s Next for Enterprise Agents

Godement offered a preview of upcoming developments. OpenAI is currently focused on:

• Multimodal agents capable of interacting through text, voice, images, and structured data
• Long-term memory to retain knowledge across interactions
• Cross-cloud orchestration to support complex, distributed IT infrastructures

These are not revolutionary changes but incremental enhancements that build on existing capabilities. "Once we have models that can process information over minutes or hours, rather than seconds… that will unlock some truly groundbreaking applications."

Final Word: Reasoning Models Are Underhyped

Godement concluded by reaffirming his belief that reasoning models—those capable of reflection before responding—will be the true drivers of long-term transformation.

"I remain convinced that we are currently at a stage comparable to GPT-2 or GPT-3 in terms of reasoning model maturity… We have only begun to explore what these models can achieve."

For enterprise leaders, the takeaway is clear: the infrastructure for agent-driven automation is now available. Success depends on identifying focused use cases, empowering cross-functional teams, and maintaining an iterative approach. The next wave of value will come not from flashy demonstrations, but from robust systems shaped by real-world needs and operational excellence.