Microsoft Agent Framework adds memory, interactive experiences, and resilient execution

Microsoft rolled out a substantial update to its Agent Framework spanning both .NET and Python, targeting the same production-readiness gap AWS just quantified. The headline additions are persistent memory that survives beyond a single conversation, interactive interfaces for richer user interaction, code-execution environments for agents to run and verify work, and resilient execution that recovers from interruptions rather than failing outright.

The framing is a shift in ambition: from agents that answer questions to agents that complete useful, multi-step work reliably. Persistent memory and interruption recovery are precisely the capabilities that separate a demo from a deployable system — an agent that forgets context between sessions or dies on a transient error isn't production-grade. Code-execution environments, meanwhile, let agents test their own outputs, a pattern increasingly seen as essential for coding and data agents.

Strategically this deepens Microsoft's enterprise agent stack, which spans Microsoft 365 Copilot, Azure/Foundry model hosting (now including GPT-6 Astra, Sol, and Luna), and the Agent Framework as the developer-facing orchestration layer. It competes directly with AWS's Strands Harness and Bedrock AgentCore, both prominent this same week, making agent frameworks a clear battleground.

The convergence is telling: AWS, Microsoft, and Google are all racing to solve agent memory, state, resilience, and observability at nearly the same moment, validating the industry consensus that these are the real blockers to enterprise adoption. The caveat is that framework maturity is proven in production, not release notes; developers will judge it on whether the resilient-execution and memory features hold up under real workloads. Watch adoption relative to open-source alternatives and AWS's tooling.