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

## Executive summary

OpenAI introduced the Agents API, offering a 'Harness-as-a-Service' built on the Codex harness. This service allows developers to build complex, long-running, agentic workflows—such as investigating outages or rebooking flights—without needing to manage the underlying infrastructure at scale. The API addresses fundamental engineering challenges, including secure cloud execution, state preservation over hours, and visualizing multi-step agent decisions.

## Key takeaways

- Managed Agentic Workflows: The Agents API provides a managed service layer over the Codex harness, allowing developers to define tools, data, and instructions while OpenAI handles the complex infrastructure required to run the agent at scale.
- Advanced Model Capabilities: Modern models like GPT-6 Astra can effectively use dozens of tools for extended periods (hours or days), necessitating a flexible foundation like the Responses API (launched in 2025) and the Agents API.
- Solving Infrastructure Complexity: The API solves critical engineering problems: securely running a local harness in the cloud, providing a useful view of the agent's decision-making process, and preserving state across long-running tasks, server deployments, or connection drops.
- New Feature: Computer Use: A new feature, 'computer use over the API,' brings the abilities of Astra to interact with a Graphical User Interface (GUI) within OpenAI-hosted environments.

## Technical details

- API Layering: OpenAI's APIs are structured in layers: the Responses API (the base layer for agentic harnesses), Codex (provides 'hands' for coding/browser use), and the Agents API (the highest layer, enabling full automation of business processes).
- Core Engineering Challenges: The Agents API addresses three key problems: 1) Moving a local harness to a secure cloud environment; 2) Providing a clear view of the agent's decision steps; and 3) Maintaining state and reliability over tasks that run for hours, even during server recycling or code deployment.
- State and Reliability Management: To ensure reliability (three nines), the system must track state precisely, allowing agents to resume tasks exactly where they left off, even if the connection drops or servers are replaced. This prevents repeating actions or losing progress.
- Multi-Agent and Tooling: The API supports multi-agent mode, allowing an agent to delegate tasks to subagents concurrently. Developers can share skills and specify paths to plugins (including MCP servers) via the `capability directories` field.

## Practical implications

- Developers can focus on business logic (e.g., rebooking rules) rather than maintaining complex distributed infrastructure.
- The service enables parallel processing of tasks, demonstrated by handling multiple passenger rebookings simultaneously.
- The API supports integration with external systems, allowing users to bring their own sandboxes or use third-party providers.
- The service facilitates building synthetic audiences (virtual focus groups) for market research, as demonstrated by Bain's use case.

## Topics

Agentic AI, Harness-as-a-Service, API Development, State Management, Distributed Computing, OpenAI Codex, Codex harness, Agents API, Responses API

Source: https://www.youtube.com/watch?v=b7l_Dhe1h8M
