Harness-as-a-Service: Building with the Agents API | DevDay 2026

Categories: AI, Product

Summary

OpenAI's new Agents API represents a fundamental shift from developers building their own agent infrastructure to a managed service model—GPT-6 Astra can now handle tasks spanning hours or days across dozens of tools, requiring OpenAI to solve three critical engineering problems: secure cloud execution, real-time observability, and state persistence across server failures.

Key Takeaways

  1. The API layering strategy follows Maslow's hierarchy: Responses API (foundation for content types), Codex (agent tools/capabilities), Agents API (managed service). This architectural progression lets developers choose their control/complexity tradeoff.
  2. Developer demand shifted from DIY infrastructure to managed services—teams knew what tools and workflows they needed but explicitly asked OpenAI to 'run it for them.' This revealed a critical gap between product vision and operational burden.
  3. Model capability expansion (GPT-4 Turbo: seconds → GPT-6 Astra: hours/days with dozens of tools) required complete software redesign around the harness, not just API wrapper improvements on existing inference stacks.
  4. Three engineering challenges defined the Agents API design: (1) separating cloud harness from execution environment, (2) providing real-time agent decision visibility during execution, (3) maintaining state across multi-hour tasks with service restarts.
  5. The Agents SDK (local agent harness from 2025) taught OpenAI that total control doesn't equal desired responsibility—developers wanted autonomy over tools/data/instructions but preferred outsourcing infrastructure operations and continuous harness improvements.

Related topics

Transcript Excerpt

Hey everyone. I'm Charlie from the developer experience team. And I'm Steve. I'm an engineer on the API team. And for well over a year, we've heard about how agents are going to change everything and be the next evolution in AI. And here in 2026, it feels like that's actually become true at the frontier. We can ask agents to investigate an outage, to solve a support ticket, or to help reconcile finances. We're seeing that shift here at OpenAI with Codex and ChatGPT work. We're using agents across the company, from engineering to finance to marketing. A big part of that, perhaps the biggest part, is the model. As models become more capable, we can ask them to take on more involved work. Astra and GPT 6.1 Sol can handle larger tasks from start to finish. But every time the model can do more,…

More from OpenAI

Featured in