The Kernel That Must Say No: how Semantic Kernel gates AI tool execution
Mahmoud_Zalt · x · 2026-09-10
AI solutions architect Mahmoud Zalt dissects Microsoft's Semantic Kernel Python kernel.py, arguing an AI orchestrator earns trust by making refusal explicit. The Kernel is not the model, plugins, embeddings or MCP server — it's the control desk in front of them: the single public interface resolving functions, selecting services, applying filters and recording outcomes. It decides which requests proceed, how failures become model feedback, and which mutable values enter ChatHistory — effectively a policy boundary. The article walks through the mixin structure (Filter/Function/Services/Reliability extensions, KernelFunction, embeddings, MCP connectors) and frames the core engineering question: where should software draw the line between proceeding and refusing?
More from coding & agent
- Full pipeline: niji + GPT Image + local MiniMax H3 + Codex for pixel-perfect animation — Hailuo_AI · 2026-09-10
- Instinct launches agent-to-agent protocol so personal AIs can coordinate your plans — mon__lim · 2026-09-10
- Traser: a local deterministic tool to triage agent runs that succeed but return wrong results — Sensitive-Parsnip-12 · 2026-09-10
- Experiment: with a false premise and stale artifacts, only one of three agents found its peers — turtle_bazon · 2026-09-10
- Iron Man fan builds interactive suit teardown site with GPT-6 Astra, Three.js — TheMoonMidas · 2026-09-10
- Hermes Agent routing flaw silently switched user to metered OpenRouter, costing $100 — WolframRvnwlf · 2026-09-10