State isolation and security boundaries
Microsoft released version python-1.21.0 of its Agent Framework on October 8, 2026. The update targets execution integrity when agents operate within hierarchical setups. When agents run as tools, the core engine, Foundry, and GitHub Copilot packages now isolate provider-owned service session state to prevent cross-call contamination.
Security mechanisms also received operational warnings. The Hyperlight and Monty integrations alert operators when CodeAct tools cannot support FIDES security enforcement. In addition, the framework added full-buffer Purview policy evaluation for streamed model responses alongside guidance for handling tool results.
Execution fixes and UI permission controls
The release addresses several tool execution and data integrity issues. The core runtime serializes shared file modifications to prevent conflicting edits and executes local sibling tools in declaration-only mixed batches. It also issues warnings whenever streamed result gates lack buffering.
For interfaces managing agent approvals, DevUI now scopes function approvals directly to their individual conversations instead of applying them broadly. In provider-specific updates, the Bedrock client retains extended-thinking reasoning traces across tool invocations so intermediate reasoning steps remain intact during multi-step execution.
Our assessment
Developers running multi-agent topologies should verify whether isolating provider session state changes memory sharing between parent coordinators and tool agents. Teams using CodeAct tools must evaluate whether lack of FIDES enforcement flags insecure execution paths in production. Administrators handling human confirmations should test DevUI to ensure scoped function approvals do not interrupt legitimate repetitive actions.
Source: Microsoft original publication.
Read the original material
Microsoft (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
