01 / behavior
Agent
The model, system prompt, tools, MCP servers, and skills that define how Codex behaves.
Managed runtime for Codex
Codex Managed Agents gives you the harness, secure tool execution, and cloud runtime for long-running agent work. Instead of building your own loop, sandbox, and orchestration stack, you get a managed environment where Codex can read files, run commands, browse the web, and execute code efficiently.
Runtime topology
Model + tools
Define Codex behavior once.
Container template
Bring packages, files, and access rules.
Execution runtime
Combine the agent and environment for a live task.
App I/O
Send instructions in and stream tool/status updates out.
Core concepts
Agent, Environment, Session, and Events form the vocabulary of the system. The homepage keeps them visible because they are the shortest path to understanding how the product fits together.
01 / behavior
The model, system prompt, tools, MCP servers, and skills that define how Codex behaves.
02 / runtime
A configured container template with pre-installed packages, mounted files, and network rules.
03 / execution
A running agent instance inside an environment, working on a concrete task and producing outputs.
04 / control plane
Messages flowing between your application and the agent, including user turns, tool results, and status updates.
How it works
The workflow is intentionally small: define Codex behavior, choose the runtime, launch a session, and keep steering it through events while the infrastructure stays managed.
Define the model, prompt, tools, MCP servers, and skills once, then reuse the agent by ID.
Provision a cloud container template with the packages, files, and access controls your workload needs.
Launch an execution that binds the chosen agent and environment for a specific task.
Push user input as events while Codex autonomously calls tools and streams back updates over SSE.
Guide the session mid-run, interrupt it, or redirect it without rebuilding the whole runtime loop.
When to use it
Codex Managed Agents is not just a UI wrapper around a model. It is a runtime layer for workloads that need time, tools, isolation, and a memory of what already happened.
Handle workloads that span minutes or hours and require multiple coordinated tool calls.
Run inside managed containers with pre-installed packages, mounted files, and explicit network access.
Skip building your own agent loop, sandbox, and tool orchestration layer from scratch.
Keep a persistent file system and conversation history across multi-turn tasks and retries.
Ready to ship
Start with the Quickstart to create your first managed session, then move into the API Reference when you are ready to wire it into your own application.