Intel / repo

openai-agents-python v0.22.1

Source: github.com
Author: seratchTop Viber
Date: Published Sep 8, 2026

From the Daily Brief

OpenAI's Agents SDK just shipped v0.22.1, and the headline change is guardrails that apply directly to MCP tool calls rather than living in application code around them. Builders can now block a tool call before it runs based on server-side policy, instead of writing their own pre-execution checks for every MCP server they connect. The release pairs that with configurable sandbox isolation and Docker sandbox labels, giving teams a way to run untrusted code with tighter environment controls baked into the framework rather than bolted on after the fact. Smaller additions round out the release: image results in web search, custom guardrail-blocked messages so a blocked call can explain itself to the user, and streamed transcription options. The SDK also fixes how tool calls and guardrails interact with max_turns, closing a gap where a blocked call could still count against a run's turn budget. None of this is flashy, but it is the kind of infrastructure work that determines whether teams trust an agent framework enough to give it access to real tools in production. For anyone building on top of MCP, upgrading gets you server-wide enforcement for free.

Read the 2026-09-08 Brief →

Context

OpenAI's Python Agents SDK adds a way to apply the same input and output checks across tools exposed through a local server connection. The feature appears in the v0.22.1 release notes. Its implementation description says the optional checks apply to every converted tool: input checks participate in the existing execution and approval process, while output checks inspect the SDK's converted result. Developers can therefore configure shared checks at the server level, with output inspection operating on the representation the framework uses.

Terms in this piece · Glossary

MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.

Read the source github.com