Summary
azure-ai-agentserver-agentframework (latest 1.0.0b17, released 2026-03-16) hard-pins both agent-framework-core and agent-framework-azure-ai to >=1.0.0rc2,<=1.0.0rc3.
This means the server adapter is incompatible with every released stable version of Microsoft Agent Framework, including 1.0.0, 1.0.1, 1.1.0, 1.1.1, and the just-released 1.2.0 (2026-04-24). Anyone wrapping a MAF agent with from_agent_framework(...) is stuck on the pre-release rc3 line.
Reproduction
python -m venv .venv && source .venv/bin/activate
pip install azure-ai-agentserver-agentframework==1.0.0b17 agent-framework-core==1.2.0
Result: pip resolves agent-framework-core back down to 1.0.0rc3 (or fails, depending on other constraints in the environment).
Verified METADATA from the wheel:
Requires-Dist: azure-ai-agentserver-core==1.0.0b17
Requires-Dist: agent-framework-azure-ai<=1.0.0rc3,>=1.0.0rc2
Requires-Dist: agent-framework-core <=1.0.0rc3,>=1.0.0rc2
Requires-Dist: opentelemetry-exporter-otlp-proto-grpc>=1.39.0
The same upper bound was present in 1.0.0b16.
Impact
- Users hosting MAF agents in Azure AI Foundry via the agent server adapter cannot adopt any of the bug fixes, features, or security hardening shipped in MAF
1.0.0 → 1.2.0 (e.g. restricted pickle deserialization in FileCheckpointStorage, hosted agent V2, Foundry Toolboxes, OpenTelemetry improvements).
- New users following the MAF "Quick Install" guide (
pip install agent-framework) will hit a resolver conflict the moment they add the server adapter.
- The pin is a
<= ceiling rather than a compatible-release range (~=), so even patch releases on the MAF side are excluded.
Ask
- Is the
<=1.0.0rc3 ceiling intentional, or a leftover from before MAF reached GA?
- If a known incompatibility exists with MAF
1.0.0+, can it be documented?
- Could a
1.0.0b18 (or GA) be cut that widens the range to something like >=1.0.0,<2 to track MAF's stable line?
Happy to test a pre-release build against our hosted agents if helpful.
Environment
azure-ai-agentserver-agentframework: 1.0.0b17agent-framework: 1.2.0 (would like to use)- Currently forced to:
agent-framework-core==1.0.0rc3, agent-framework-azure-ai==1.0.0rc3
- Python: 3.10+
- OS: macOS / Linux containers (Foundry hosted agent runtime)
Summary
azure-ai-agentserver-agentframework(latest1.0.0b17, released 2026-03-16) hard-pins bothagent-framework-coreandagent-framework-azure-aito>=1.0.0rc2,<=1.0.0rc3.This means the server adapter is incompatible with every released stable version of Microsoft Agent Framework, including
1.0.0,1.0.1,1.1.0,1.1.1, and the just-released1.2.0(2026-04-24). Anyone wrapping a MAF agent withfrom_agent_framework(...)is stuck on the pre-releaserc3line.Reproduction
Result: pip resolves
agent-framework-coreback down to1.0.0rc3(or fails, depending on other constraints in the environment).Verified
METADATAfrom the wheel:The same upper bound was present in
1.0.0b16.Impact
1.0.0→1.2.0(e.g. restricted pickle deserialization inFileCheckpointStorage, hosted agent V2, Foundry Toolboxes, OpenTelemetry improvements).pip install agent-framework) will hit a resolver conflict the moment they add the server adapter.<=ceiling rather than a compatible-release range (~=), so even patch releases on the MAF side are excluded.Ask
<=1.0.0rc3ceiling intentional, or a leftover from before MAF reached GA?1.0.0+, can it be documented?1.0.0b18(or GA) be cut that widens the range to something like>=1.0.0,<2to track MAF's stable line?Happy to test a pre-release build against our hosted agents if helpful.
Environment
azure-ai-agentserver-agentframework:1.0.0b17agent-framework:1.2.0(would like to use)agent-framework-core==1.0.0rc3,agent-framework-azure-ai==1.0.0rc3