AI
How a Contrarian Research Bet Became Phi
When the industry was racing toward larger parameters and more data, a small Microsoft Research team asked how much intelligence could fit in a small package. The answer: focus on data quality over quantity and use GPT-series models to generate textbook-like training data. One researcher got the idea reading bedtime stories to his daughter.
That project became Phi. Phi-2 debuted at Ignite; by Build 2025 the fourth generation runs as Silica on every Copilot PC and powers Japan Airlines’ offline reporting at altitude. Phi-4 Reasoning is now in Foundry Labs.
Azure AI Foundry Labs as a Developer Research Feed
Azure AI Foundry Labs launched last year as a front door to early-stage Microsoft Research output. Without it, these models sit in GitHub repos. Labs curates them into a single catalog developers can browse and test. The catalog includes OmniParser for UI parsing, EvoDiff for protein structure generation (thousands of structures per second), and Muse for game simulation.
Labs is where you experiment; Azure AI Foundry is where you deploy. A subset of Labs models is already available as deployable collections in Foundry, and that subset grows based on developer feedback.
Cutting Prototype Adoption from Days to Minutes with MCP
Adopting OmniParser the traditional way would have taken Shannon Monroe’s team five days: read the research, pull the repo, work through the docs. They built an MCP server loaded with the repo and documentation instead, and integration time dropped to 30 minutes.
“It became faster to automate the task than to perform the task.” — Shannon Monroe
The OmniParser cognitive load demo (19:52) uploads a screenshot and returns a heat map of where users will struggle. Adoption cycles fell from 30 minutes to 10, then to a record of 3 minutes and 44 seconds. The MCP server template is open on GitHub and works with GitHub Copilot in agent mode.
Multi-Agent Teams, AutoGen, and the Human Oversight Layer
AutoGen, released 18 months ago, decomposes tasks across specialized agents. Magentic One sits on top: a coder, executor, file browser, web surfer, and an orchestrator that maintains a task ledger to track progress and recover from failures. In the Magentic One stock analysis demo (35:08) the orchestrator delegates subtasks and delivers a final answer. That team hit number one on multiple agentic benchmarks.
“Communication is not the same thing as collaboration.” — Ece Kamar
Magentic-UI, released Monday at Build, adds the human layer. Actions run inside Docker containers. The agent presents a plan before executing; irreversible actions (ordering, emailing) require explicit approval, and the decision is stored for next time.
Q&A
OmniParser works by screenshotting everything — will there be a protocol letting agents see what a system can do directly, like AppleScript dictionaries? Ece Kamar described NL2Web as the structured alternative: it converts any website into a callable API, but OmniParser remains a fallback because content providers may block API-style access to protect ad revenue. Both approaches are being pursued in parallel. ▶ 50:13
How should I choose between AutoGen, Semantic Kernel, and Magentic One for client work? Ece Kamar said the confusion will resolve: both AutoGen and Semantic Kernel are being merged into a single Agent Service stack inside Foundry, with Magentic One sitting on top as a deployable generalist agent team rather than a platform library. ▶ 54:05
Are we repeating the Internet’s identity and access management mistakes with agentic AI? Vaidya said the Agent Service handles enterprise auth and function-call accounting today; Ece added that what agents should and should never do is still an open research question, citing a Magentic One experiment where the agent independently started emailing a government agency to request missing public data. ▶ 55:53
Notable Quotes
It became faster to automate the task than to perform the task. Shannon Monroe · ▶ 19:12
Communication is not the same thing as collaboration. Ece Kamar · ▶ 32:24
it is clear that the future is multi-agent. Ece Kamar · ▶ 31:58
Key Takeaways
- Foundry Labs gives developers direct API access to pre-production Microsoft Research models before they ship.
- Magentic One pairs a task-ledger orchestrator with specialized agents to hit number one on multiple agentic benchmarks.
- Magentic-UI adds co-planning and action guards so humans approve irreversible steps before agents execute them.