MICROSOFT JUST OPEN SOURCED THE ENTIRE AGENT RUNTIME, SO YOU NEVER HAVE TO BUILD ONE AGAIN a raw model only writes text. to make it a real agent it needs a loop, tools, memory, planning, approvals, and a way to know when it is done. Microsoft's Agent Harness is... that whole layer, and it ships with everything on by default. you bring the model and the tools, it wires up the rest: > an automatic tool calling loop with iteration limits > history saved after every call, so a crash resumes instead of starting over > context compaction that summarizes when the window fills > a persistent todo list plus plan and execute modes > file memory that survives across turns > tool approval with standing rules and auto approval > OpenTelemetry tracing and web search built in > Agent Skills that load domain expertise on demand the mental model is simple: the model is the brain, the harness is the body around it. and it works with whatever you already run: > Azure OpenAI, OpenAI, Anthropic, and GitHub Copilot, and it hosts on Microsoft Foundry you spin one up in a single line: agent = create_harness_agent(client, name, instructions) open source, Python, .NET, and Go, stable since mid 2026. bookmark this. Microsoft already built the hard part of every AI agent, you just plug in the model.show more

microsoft's agent harness is a bandage on a broken architecture. it ends with brittle agents failing in production by late 2027.

The hard part was never the harness, it's catching what slips through your own PRs. I once filed a bug report against two of our own GitHub PRs, one never merged and one where changes looked erased after closing, plus broken cross-tab navigation on a staging site.

One line to spin up an agent, wild 🤯

Great share

Agent Harness is the interesting layer here. The Harness Is The Product. The Model Never Was

Demos skip approvals and persistence. Real work does not. That boring machinery is why the runtime still works on Tuesday.

This is another strong example of the industry moving toward agent infrastructure instead of just better models. The important distinction is: Model = intelligence Harness = operational system around that intelligence A production agent needs much more than an LLM: tool execution, memory, state recovery, planning, permissions, context management, observability, and termination logic. Microsoft’s approach packages those pieces so developers don’t have to rebuild the same agent loop repeatedly. Where VOOL fits This is highly relevant to VOOL because it reinforces the same architectural direction: agents need a governed execution environment around the model. VOOL’s role is particularly focused on the control layer—policy-shaped model routing, explicit tool contracts, scoped approvals, local-first execution, and separation between planning, tool execution, and payment actions. So the larger stack can be thought of as: Model → Agent Harness → Governance → Tools → External Systems Microsoft’s harness primarily addresses how an agent operates continuously. VOOL’s architecture addresses how that operation is governed—which model should run, which tools are available, what actions require approval, and how execution is constrained. And that’s an important distinction as agents become more autonomous: Giving an AI a body makes it capable. Giving that body rules makes it deployable.

the hard part will be debugging state and approval behavior once production workflows get messy.

finally, my local repo can stop looking like a graveyard of abandoned state machine attempts.

That’s so huge bro

Microsoft is late to the party.

the missing loop layer matters tools memory approvals all shipped

crash resume and standing approvals stay after you swap brains

The runtime loop is the easy-to-demo part. I’d want the harness to make “done” testable: which source state, which permissions, and what happens when another worker changes the row mid-run?

approvals as a feature. who actually hits it at 6am when the thing's already moving.

The loop is the easy half. Approvals and "when is it done" are what break in real use. I'd test one thing first: give it a task with no clear finish line and see if it stops or spins for 20 minutes.