OpenAI's Agents SDK now blocks empty tool arguments by default

OpenAI's Agents SDK 0.22.1 changes how the framework handles a tool call when argument parsing fails. Before this release, empty or malformed tool arguments still let the tool run using its defaults. Now that path fails closed instead.

That matters for anyone building agents that call tools with real side effects. A tool that runs on defaults because its arguments didn't parse is a bug that passes testing. It only shows up in production, on a request nobody watched closely enough to catch.

The release also adds server-wide guardrails for MCP tools and configurable Unix-local sandbox isolation. Docker sandboxes get container labels, web search results can include images, and blocked-output messages are now customizable instead of fixed text.

Two more fixes sit in the same argument-handling area as the empty-arguments change. The SDK now rejects **kwargs keys that collide with a tool's named parameters. It also recovers failed resumed Session writes before making the next model call. Both close gaps where a tool could act on data that didn't match what the caller intended.