THE OPENAI AGENT ECOSYSTEM, EXPLAINED

Understand the stack.
Build better agents.

Explore how the Agents API, models, tools, and infrastructure connect. From your first session to production.

Explore the agent ecosystem

Ecosystem mapClick a component to explore
tasks ↓   ↑ events
Sessions · Context · Orchestration
inference
tool calls
executes in
results & saved work ↓
Solid connections show dependenciesEnvironment is optional
Component overview

Agents API

OpenAI runs the Codex harness: the model-and-tool loop, durable sessions, context compaction, and recovery.

UPSTREAM
Application requests and configuration
DOWNSTREAM
Models, tools, optional environments
Explore this topic
Based on OpenAI documentation · Reviewed September 12, 2026View sources

A clear path through the ecosystem

Start with the big picture, choose your runtime, or go straight to code.

What happens after you send a task?

Walk through the request, tool call, and result one event at a time.

ILLUSTRATIVE SESSION FLOW

Submit a task

Your server creates a session and supplies instructions, input, and permitted tools.

Application → Agents API

Go deeper into the stack

Common questions, clear answers.

Is the Agents API the same as the Agents SDK?+

No. The Agents API runs a managed Codex harness at OpenAI. The SDK runs its agent loop in your application runtime. They have different resource and deployment models.

Do I need a sandbox for every agent?+

No. Use environment.type: none for tasks that can use remote MCP and application functions without a filesystem or command execution.

Can I use my own infrastructure?+

Yes. A self-hosted environment connects your compute to the managed harness through an executor. You own that environment’s provisioning and lifecycle.

Is this an official OpenAI website?+

No. Agents API Encyclopedia is an independent educational resource. Each topic links to the relevant official documentation.