Middleware architecture overview

What is a middleware in the ReAct loop?

A hook into the ReAct loop — e.g. "before calling the model," "after a tool," "before responding to the user."

What does Code Interpreter middleware do?

The model writes code, and the code runs in an isolated environment.

What does Sandbox middleware do?

Provides an isolated environment so the agent can't break your system.

What does Filesystem middleware do?

Provides file access: reading, writing, document analysis.

What does Planning middleware do?

The agent builds a plan (step 1 → step 2 → step 3) and then executes it.

What does Subagents middleware do?

Spawns child agents for parallel work.

What does Human-in-the-loop middleware do?

Before an important action, the agent asks for permission.

What does Skills middleware provide?

Reusable skills, i.e. pre-trained patterns.

What does Memory middleware provide?

Long-term memory across sessions.

What does Summarization middleware do?

Compresses long history to fit the context window.

What does Context offloading do?

Offloads part of the context to external storage and loads it back as needed.

What does Prompt caching do?

Caches part of the prompt so it isn't recomputed every time.

What is an agent without middleware?

Just: request → model → response.