Microsoft Agent Framework vs Google Agent Development Kit (ADK)

Microsoft Agent Framework (MAF) and Google Agent Development Kit (ADK) answer the same question from two different hyperscalers: should a major AI lab ship its own official, opinionated agent SDK instead of leaving the category to third parties like LangChain or CrewAI? Both say yes, but they arrived from opposite directions. MAF is explicitly a consolidation. Microsoft merged two previously separate stacks, AutoGen's multi-agent orchestration and Semantic Kernel's enterprise plumbing, into one supported framework, and put both predecessors into maintenance mode (bug and security fixes only) to push new development toward it. It reached a production-ready 1.0 GA on 2026-04-03. ADK is a from-scratch product: Google announced it at Cloud NEXT in April 2025, a full year earlier, and built it around protocol interoperability and evaluation tooling from day one rather than merging legacy stacks.

Checked live via the GitHub API on 2026-08-06: Microsoft's repository (microsoft/agent-framework) carries 12,622 stars and 2,117 forks; Google's flagship repository (google/adk-python) carries 21,013 stars and 3,803 forks. Both were pushed to within the last day of this research, real active projects rather than abandoned ones. Language spread is the clearest structural difference between them. MAF ships production-grade SDKs in three languages (.NET, Python and Go), reflecting its .NET/Azure-enterprise roots; ADK ships four (Python, Go, Java and TypeScript, all reaching 1.0 GA) plus an experimental Kotlin/Android runtime, the broadest spread of any framework in this index. A team standardizing on Java or TypeScript has a first-party GA SDK on ADK and none on MAF.

Both ship native support for the two dominant 2026 agent-interoperability standards, the Agent2Agent (A2A) protocol and the Model Context Protocol (MCP), so agents built on either framework can in principle hand work to agents built on the other. Where they diverge is which production concern each treats as first-class. MAF's Agent Harness gives an agent guarded shell and filesystem access, human-in-the-loop approval flows and automatic context compaction; its CodeAct mode, where the model writes a short program to batch tool calls inside an isolated micro-VM instead of calling them one at a time, is the one benchmarked performance claim in this comparison, with Microsoft's own BUILD 2026 numbers reporting 52.4% faster completion and 63.9% fewer tokens on multi-step workloads. ADK's built-in strength is evaluation: `adk eval` scores an agent's full trajectory and final output against repeatable test cases, and a local ADK Web UI debugs a flow before shipping it, a stronger built-in testing story than MAF ships today, though ADK has no equivalent to the Agent Harness's guarded real-environment access.

Pricing is free-versus-free at the open-source core: MAF is MIT-licensed, ADK is Apache 2.0, and both are fully self-hostable at no cost beyond the model provider's own bill. The managed layers price differently. MAF's Foundry Hosted Agents publishes no flat fee at all; it bills purely as ordinary Azure/Microsoft Foundry consumption, so there is no separate number to compare. ADK's Agent Runtime instead bills separately for compute, memory and session storage, with published monthly free allowances for compute and memory. Neither managed layer is required to use the framework; both cores run anywhere for free.