OpenAI's Agents API is a managed framework that runs agentic AI workflows on your behalf. You define the agent (model + instructions + tools), pick an execution environment (OpenAI-hosted sandbox or self-hosted), and then create sessions where the agent completes tasks across multiple turns while OpenAI handles session state, context windowing, and orchestration. The system bills per model call plus sandbox costs. The practical angle: you're trading control and customization for durability—sessions persist, so you can hand off work to an agent, come back to it, steer mid-task, or spin up subagents for parallel work. The source lists concrete examples (incident response, data analysis, GitHub debugging) and shows the API currently supports MCP server connections, web search, and programmatic tool calling. One thing missing: whether there are guardrails on what agents can actually do, or if sandboxing alone prevents drift.