Fetching from the wire…
Public story · 2026-08-05 · high
Microsoft's new Agent Framework harness routes every shell command and file write through a permission handler you write yourself, and a related update stops silently swallowing agent errors.
Why now: Microsoft published the harness and the Agent Framework 1.17.0 changes on August 4.
Microsoft shipped the GitHub Copilot Agent for Agent Framework on August 4, and the detail that matters is what it doesn't let you skip. Every shell command, file write, URL fetch, and custom tool call routes through a permission handler you have to define yourself. There's no default-allow path. The package lands as Microsoft.Agents.AI.GitHub.Copilot for .NET 8+ and agent-framework-github-copilot for Python 3.11+, per Microsoft's announcement.
This matters because most agent harnesses treat approval as a setting you turn on if you're careful. Microsoft built it as the only path through. If you're wiring an agent to touch a filesystem or hit a network endpoint, you write the gate before the agent runs, not after something goes wrong.
The same release carries a second change that's easy to miss and probably more disruptive. Agent Framework 1.17.0 pulled Durable Task and Azure Functions integrations out of the core packages, so anyone depending on them directly now needs a separate reference. More significant: declarative workflows now fail when an agent returns an error, instead of continuing silently. Per Microsoft's post, that's a direct behavior change, not just a bug fix.
If your workflow has been running clean for months, that may be because a step has been failing and the framework kept going anyway. Upgrading to 1.17.0 turns that into a visible failure. Worth checking your error logs against what actually ran before you assume anything's been passing that actually wasn't.
The harness change and the workflow change point the same direction: Microsoft is closing gaps where agent behavior went unchecked by default. Whether that's the whole story for this release, Microsoft's post doesn't say.
Each link below shares sources, entities, or timing with this story.
Microsoft's update ships a new Agent (Preview) in Copilot Chat built on the SDK powering Copilot CLI, explicitly so behavior stays consistent across CLI, the GitHub app, VS Code and Visual Studio (Visual Studio Blog). It ships curated .NET and Azure skills you enable selective...
July 23, agent-framework-declarative went 1.0.0 for Python, matching the already-stable .NET package (Microsoft). Workflows author in YAML with Power Fx formulas for state and expressions, conditions, loops and jumps, sequential and conditional routing, function tools, MCP and...
InfoQ reported the Agent Framework Harness and Foundry Hosted Agents reaching GA in .NET and Python on August 3, bundling function invocation, per-call history persistence, context compaction, todo lists with plan/execute modes, file memory, skills, web search, tool approval,...
Microsoft dropped the [Experimental] tag and shipped a production Agent Skills API in Agent Framework. Agents load reusable packages of instructions, docs, and scripts only when a task needs them, via load_skill, read_skill_resource, and run_skill_script, and all three require...
A spec is a press release until someone who didn't write it implements it. GitHub made Agent Plugins 1.0 generally available on August 12 across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app on all plans. The spec, published August 6, was co-authored by AWS, Anysp...
Microsoft's Agent Framework blog confirms the MCP 2026-07-28 revision is in release candidate with a final lock on July 28, while A2A passed 150 adopting orgs with production SDKs in Python, JS, Java, Go, and .NET. The next two weeks are your window to review MCP breaking chan...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.