Microsoft Agent Framework blocks workflow reuse across authorization scopes
Microsoft Agent Framework’s Python release assigns each active workflow executor and session to a single authorization scope. Reuse from another scope stops before the model is called. The November 7 update also limits external-tool discovery and preserves a complete catalogue when refresh fails. Migration requires fresh workflows and sessions for each call, alongside a newer Python environment.
Artificial Intelligence··Midday
Live workflows are bound to the caller’s scope
Microsoft Agent Framework, a toolkit for building AI agent workflows, changed authorization handling in its November 7 Python 1.25.0 release. Live executors and sessions belong to one authorization scope. Reusing them in another scope is rejected before any model call. The migration guidance calls for newly created workflows, executors and sessions for each invocation. The check covers live identities, with shared-storage state and custom functions outside its stated protection.[1]
Tool discovery stops at bounded catalogue limits
For external tools using the Model Context Protocol, a common agent-tool connection method, catalogue discovery is capped by default at 1,000 entries across 100 pages. Framework-managed requests normally time out after 30 seconds. If refreshing a catalogue fails, the previous complete snapshot stays available. Larger catalogues can use increased limits, while caller-supplied clients retain their own connection and timeout policies.[1]
Migration also replaces implicit session continuity
For every invocation, declarative workflows now create a protocol session and close it afterwards. Using cache_max_size, the former setting, produces a warning and has no effect. Applications requiring continuity can use explicit remote-state handles. Installing this release also requires Python 3.11 or later, replacing eligibility for Python 3.10.[1]
