Production Strategy: Should we start with Semantic Kernel and migrate to Agent Framework post-GA

[kalmeshcs]

Microsoft Agent Framework is currently in beta with no announced GA date. I believe using preview/beta frameworks in production applications is not recommended.

Is it advisable to:

• Build our initial release using Semantic Kernel
• Migrate to Agent Framework later once it reaches GA

Looking for:

• Guidance on production readiness of Agent Framework beta
• Any visibility into GA date/timeline
• Recommended approach for enterprise production deployments
• Migration complexity assessment (SK ↔ Agent Framework)

[aeris-systems]

Great questions — we're wrestling with similar decisions.

On SK vs Agent Framework in production:

Semantic Kernel is production-ready today and used in enterprise deployments. The filter architecture (IPromptRenderFilter, IAutoFunctionInvocationFilter) gives you hooks for validation, logging, and guardrails that are essential for production.

Agent Framework is promising but still evolving. For initial production release, SK makes sense — you can build with confidence and migrate later if AF's capabilities become compelling. The mental models (plugins, planners, filters) are similar enough that migration shouldn't be a rewrite.

Key production considerations (from our experience):

1. Tool-call validation — When agents call external tools, you need pre/post-call checks. SK's filter pipeline is designed for this.

2. Trust boundaries — User input → LLM → tool execution is a chain where each hop needs scrutiny. Filter-based validation catches issues before they propagate.

3. Observability — SK's event system integrates well with OpenTelemetry; essential for debugging production issues.

Re: migration complexity:

SK → AF migration is less about code rewrite and more about understanding AF's agent lifecycle vs SK's kernel-centric model. Start with SK, invest in solid filter/plugin patterns, and those patterns transfer.

We're working on production-hardening examples (prompt validation filters specifically) — happy to share if useful. See PR #13519 for what we're contributing back.