python-1.18.0
Microsoft Agent Framework 1.18.0 introduces shared vector-store abstractions with multiple provider implementations, while adding breaking changes to SecretString, checkpoint deserialization, dependency boundaries, and file-store contracts.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
Microsoft Agent Framework release python-1.18.0 delivers several architectural updates, new vector store connectors, and multiple breaking changes: Breaking Changes: - SecretString: Changed from a str subclass to a masked value wrapper across core and provider integrations (Anthropic, Azure AI Search, Cosmos, Bedrock, Copilot Studio, Foundry, Gemini, Mem0, OpenAI), aligning credential handling. - Dependencies: Isolated Lab dependency resolution, removed Lab from core 'all', and updated Azure AI Projects support through 2.6 with OpenAI 3.x. - agent-framework-foundry-hosting: Checkpoint deserialization is restricted by default; applications must register custom checkpoint types. - agent-framework-github-copilot: Workspace file hooks are now opt-in. - agent-framework-core (Experimental): Centralized path normalization for files, memory, sessions, and skills. Added ranged file reads and shifted line-numbering contracts onto AgentFileStore. New Features and Additions: - Vector Stores: Introduced shared vector-store abstractions and portable filters with implementations for In-Memory, Azure AI Search, Redis (HASH and JSON), Qdrant (alpha), and PostgreSQL/pgvector (alpha). - Execution Controls: Added maximum-duration bounds and stop-reason signals to the core tool invocation loop. - Workflows: Added support for mixed workflow invocation keyword arguments. - UI & Interop: AG-UI now supports MCP Host-history conversion for persisted conversations and configuration to suppress terminal message snapshots. - Integrations: Updated Hyperlight Sandbox to 0.6 with Python 3.14 support,…
可执行摘要
Agent Framework 1.18.0 introduces breaking changes across multiple packages: SecretString is now a masked value wrapper rather than a str subclass, checkpoint deserialization is restricted by default in foundry-hosting, workspace file hooks in GitHub Copilot are opt-in, and Lab is removed from core all. Additionally, it adds shared vector store abstractions (in-memory, Azure AI Search, Redis, Qdrant alpha,…
- Agent 实用度
- 82/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Update code handling SecretString to accommodate its change from a str subclass to a masked wrapper.
- Explicitly enable workspace file hooks if using agent-framework-github-copilot.
- Register allowed checkpoint types in agent-framework-foundry-hosting to prevent deserialization errors under the new restricted default.
- Update import and dependency configurations if relying on Lab being part of core 'all'.