Microsoft Agent Framework: build your own 'claw' agent harness

Microsoft's Agent Framework team kicked off a tutorial series teaching developers to build their own 'claw' — a CLI-style agent harness that loops around a language model wired up with tools, planning, and memory. Part one stands up the basic loop for a personal-finance assistant, walking through how the harness orchestrates model calls, tool use, and state between turns.

The pedagogical framing — 'build your own agent harness' rather than 'use ours' — is notable. It reflects a maturing developer understanding that agents are fundamentally a loop plus tools plus memory, demystifying what frameworks like LangChain, Claude Code, or GitHub Copilot CLI do under the hood. For Microsoft, teaching the primitives also funnels developers toward its Agent Framework and Azure tooling.

The content lands the same week Microsoft is consolidating internal AI tooling around GitHub Copilot CLI and away from Claude Code, so a public 'build your own harness' series doubles as positioning Microsoft as the place to learn agent engineering. It also dovetails with the broader 'agentic AI is the new computing paradigm' sentiment on Hacker News, where AgentCore, agent-loop latency, and multi-model routing are hot topics. Watch how the series progresses into harder territory — planning, long-running memory, and verification — where production agents actually break.