← Back to hub
Middleware — the 1.0 control layer
In one line
Plugins that wrap the model call and change behavior without rewriting the agent.
Ready-made middleware
| Middleware | What it does |
|---|---|
PIIMiddleware | Redacts or blocks private data. |
SummarizationMiddleware | Shrinks long chat history. |
HumanInTheLoopMiddleware | Pauses for approval on risky tools. |
Custom: subclass AgentMiddleware and hook wrap_model_call — e.g. pick a bigger model for expert users.
New: RubricMiddleware (Deep Agents)
Added June 2026 (beta). It makes an agent grade its own work and try again until it is good enough. You set up the middleware once, then pass a rubric (a checklist) when you call the agent. A small grader sub-agent scores each item. If any item fails, the feedback goes back to the agent and it runs again — until the rubric passes or it hits max_iterations. Best for tasks with clear pass/fail rules, like "all tests pass" or "every section is covered".
from deepagents import RubricMiddleware, create_deep_agent
grader = RubricMiddleware(
model="anthropic:claude-haiku-4-5", # small, cheap grader
system_prompt="Grade the code against the rubric.",
tools=[run_test_suite],
max_iterations=5,
)
agent = create_deep_agent(model="anthropic:claude-sonnet-4-6", middleware=[grader])
result = agent.invoke({"messages": [...], "rubric": "- All tests pass\n"})Note. Beta — the API may still change. If you pass no
rubric, the middleware does nothing.Links
My notes
Add your own findings here as you learn.