← Back to the briefs
Tools

Microsoft tightens MCP approvals and adds in-run compaction in agent-framework 1.24.0

Microsoft Agent Framework python-1.24.0 introduces default approval settings for local MCP tools, adds experimental strategies for agent-triggered compaction and skill unloading, and enforces stricter checkpoint deserialization checks.

Tool approvals and declarative enforcement

Microsoft Agent Framework python-1.24.0 introduces an explicit default approval mode for local Model Context Protocol tools and prompts. This baseline applies whenever tools fall outside explicit per-tool lists. The core framework also warns developers when configuring options that leave fallback approval implicit.

The update reinforces execution safety in the declarative subsystem. Registered function approval rules remain mandatory during runtime execution even if an invoking action tries to omit or disable approval requirements.

Context compaction and state security

To help manage memory consumption during long multi-turn sessions, the core module adds experimental mechanisms for agent-driven context reduction. Implementations can use AgentTriggeredCompactionStrategy and UnloadSkillsCompactionStrategy to prune context and unload unused skills during an active execution run.

State preservation receives defensive hardening as well. The core runtime rejects metaclasses during restricted checkpoint reconstruction, blocking aliases and explicit allowlist matches while keeping ordinary application types intact.

Breaking storage changes and runtime baselines

Teams deploying beta Redis context storage face a manual migration step. The provider enforces exact scope identity and refuses legacy raw-scope schemas. Setting overwrite_index to true will not convert older indexes automatically, requiring operators to wipe existing context hashes and re-index stored context data.

Microsoft has established Python 3.11 or later as the standard supported baseline. While Python 3.10 installs continue to function without guaranteed compatibility in this build, the maintainers plan to remove Python 3.10 support entirely in the next framework release.

Our assessment

Teams building stateful agents should test whether in-run compaction strategies reliably reduce token footprints without dropping critical prompt instructions. Developers operating Redis memory stores need to schedule downtime to delete outdated indexes and re-index their data before adopting this release. Projects still anchored to Python 3.10 environments must upgrade their runtime base immediately before the next release cuts off package installation.

Source: Microsoft original publication.

Original source

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.

Search the publication

Find a story by title, topic, or keyword.

Press Escape to close

Ask HorizonYour site guide
Your guide to the publication

Ask Horizon.

Ask a question about the publication.

Privacy