Microsoft AI Frontiers

Hey AutoGen community — we have an important update to share! You know us as the team behind AutoGen, one of the most popular open-source frameworks for building multi-agent systems. What you might not know is that AutoGen came out of MSR AI Frontiers, a boutique lab inside Microsoft Research. AutoGen has since graduated into the Microsoft Agent Framework, where it continues to grow with a broader team. Meanwhile, our lab has kept pushing on the frontier of agents and the models that power them: small models that punch above their weight (Phi-4 Reasoning, Fara-7B), powerful agents that work across the browser and terminal (Magentic-One, Magentic-UI), and new ideas about how models think, reason, and act. We're repurposing this account as the home for MSR AI Frontiers. Same team, still shipping at the bleeding edge of agents, with a lot more to share soon. Follow along!

And read about more in our blog post: microsoft.com/en-us/research…

Most web agents drive a browser one click at a time. We tried something different and it worked better than we expected. Webwright, a new project from our team, gives the model a terminal instead of a click loop. The agent writes Playwright code, spawns browser sessions on demand, and ends with a reusable script rather than a transient session. The results: SOTA on long horizon web benchmark Odysseys (60.8%, a 16-point jump over the previous best) and 86.7% on Online-Mind2Web with GPT-5.4 — the highest of any open-source AutoEval recipe we know of. All from a minimal harness that's roughly 1K lines of code with no multi-agent orchestration. The broader bet: as models get better at code, the right harness gets smaller, not larger. Great work by @Adamlu28 @Xu_Lingrui_ @huang_chao4969 @ahmed @AhmedHAwadallah You can check it out: microsoft.github.io/Webwright/

Safe agents don’t guarantee a safe ecosystem of interconnected agents. Microsoft Research examines what breaks when AI agents interact and why network-level risks require new approaches. Learn more: microsoft.com/en-us/research…

The unsettling part: none of these failures show up in single-agent evals. They only emerge when agents interact. The hopeful part: a few agents spontaneously developed security instincts — one even wrote a privacy manifesto that other agents started adopting as norms. The takeaway: if you're connecting agents with different identities together, you need network-level red teaming, not just per-agent benchmarks. We're working on new benchmarks to measure these network-level risks. Stay tuned for more!

One example that stuck with us: 🪱 One message. That's all it took. A single "relay game" prompt hopped agent to agent, extracting private wallet data at every stop. It looped back to the start and kept going for 12+ minutes, burning through 100+ LLM calls billed to the victims.

None of us is as dangerous as all of us. We just explored what happens when "safe" AI agents start talking to each other at scale. Spoiler: things get weird fast. 🧵👇microsoft.com/en-us/research…

Tune in on May 14 👀

Can't wait to see what you all build!

🚀🚀 Earlier this summer our work on @pyautogen graduated to product & now @Microsoft has combined AutoGen and Semantic Kernel. What this means for agent developers is one single SDK that combines the best of both worlds and learnings. Sp congrats to @ekzhu and @vykthur!

The Semantic Kernel and @pyautogen team have worked hard over the last few months to bring together the best of both frameworks for a much stronger developer experience (and to simplify the choice!). Today we are excited to release Agent Framework in public preview and get feedback. Agent Framework brings the production ready focus of Semantic Kernel with full C#/Python support, as well as the nimble intuitive interface of AutoGen for rapid experimentation. Some of the things that I think we have done really well: - Middleware support - write hooks to control agent behaviour great for RAI, security, context engineering etc - Support for Agents backed by various services - Azure Foundry Agent, OpenAI Responses, Google A2A agents, ofcourse with full MCP support - Bring your own model, integrate with any agent - Workflow API that lets you express business or application logic as a computation graph (this is something that was a weak spot for AutoGen) Ofcourse, frameworks are only there to simplify aspects of your work - what matters is what you build! To illustrate this, I’d try to do a “demo a day” over the next week to showcase capabilities of the framework. The video below shows an sample app (DevUI) that ships with Agent Framework for running and debugging your Agent Framework tools. If you are starting a new AI Agent project - try it out!

Interested in the technical details behind Magentic-UI? Our report is now available on arXiv: arxiv.org/abs/2507.22358

🚀 AutoGen v0.6.4 is out! Shout-out to @github Copilot for helping author these new features! 🧠 GraphFlow now retains execution state after termination, just like other group chats. Resets only when the graph fully completes. ⚙️ New parameter_override in Workbenches for smarter client-side tool optimization. 🔗github.com/microsoft/auto…

One of the best decision we've made early on was to NOT use layers of abstraction and go native clients route instead. The second best decision was to use @pyautogen to build all of our workflows on top of. It hits all the right pain points running in distributed environments without adding too much unnecessary bloat that makes most other solutions practically unusable. The code has never been nicer!

Check out how to configure an MCP Agents in Magentic-UI here: #mcp-server-configuration" target="_blank" rel="nofollow noopener">github.com/microsoft/mage…

🚀 Introducing MCP Agents in Magentic-UI! Spin up custom agents that wrap one (or many) MCP tools, and let the Orchestrator pick the best agent for every step of the plan. Check out the demo below to see them in action 👇 #MCP #MagenticUI #AIagents

Curious how to build agents that can control a browser? I just wrote up a full tutorial on how to do it completely from scratch and with Magentic-UI. My goal is to demystify browser-use and CUA agents, it's fun to follow along! Link: husseinmozannar.github.io/#/blog/web_age… Jupyter notebook: github.com/microsoft/mage… 1/4

My recent talk on challenges in developing human-centered agents is now available online! It provides an HCI perspective of our learning from developing @pyautogen youtube.com/watch?v=O5jSX8…