harness-selection

Build the agent loop — it is genuinely 50–150 lines, and every framework's loop is a thin wrapper over the same logic plus opinions you did not choose. Adopt only thin substrates like vendor HTTP SDKs and durable execution behind owned interfaces.

Apply the dependency test: if the dependency sees prompts and tokens, it will churn with model APIs and you should own that layer. If it sits behind the loop without touching prompts, like Temporal or sandboxing, adopt it as ordinary infrastructure.

No. Both AutoGen and Semantic Kernel are in maintenance mode; AutoGen is explicitly community-managed with no new features. Do not start anything new on either — Microsoft Agent Framework is the successor, mainly for .NET shops accepting Azure lock-in.

The SDK wrapper source is MIT-licensed, but it bundles the Claude Code CLI, which is proprietary under Anthropic's Commercial Terms. Use it inside your runtime with operator-provided entitlements, but never vendor it into a generated artifact — declare it as a peer dependency instead.

Adopt durable execution only when a declared durability constraint demands it — never build it yourself. Four independent teams (Pydantic AI, LlamaIndex, Mastra, Microsoft Agent Framework) integrated existing durable execution rather than building their own.

Pre-1.0 frameworks churn rapidly: the two lab SDKs with the most agent traffic remained pre-1.0 after 18+ months, and frameworks like AutoGen entered maintenance mode within two years. Generated artifacts bound to them break on someone else's release schedule.