Source
By builders, for builders.
Today, we’re introducing run-assert-eval, a skill that discovers the risks that matter for a given agent, measures how often the agent fails, generates runtime policy directly from those findings, and reruns the eval to prove whether the fix worked.
Learn how the Python Risk Identification Tool for generative AI evolved over time to open up AI red teaming to subject matter experts from virtually any domain.
Shared design components aren’t enough to ensure a seamless product experience. Can designers create consistency by embracing simplicity over complexity?
Agents are moving into production faster than the governance around them. Today’s controls are framework-specific, mostly observe-only, and fail open when they crash. To help address this, we created Agent Hooks: an open, framework-neutral governance contract with conformance testing on both sides. What follows is the contract and the story of proving “deny means deny” inside a real framework core.
A safety requirement might appear simple while the runtime failure surface is complex. ASSERT helps identify coverage bugs; ACS fixes the policy.
Human review improves safety but limits autonomy. Standing credentials preserve autonomy but increase risk. With Azure SRE Agent, we found a safer middle by moving control out of the model and into the runtime around it.
ThinkingBox separates the execution framework from the benchmark package. This split lets builders update the harness and benchmark independently.
Following the release of TypeScript 7.0, Hejlsberg joins host Ryan Peterman to discuss the process of porting from JavaScript to Go, the pros and cons of agentic AI for coding, and more.
Budget caps and gateways can’t stop a runaway AI agent mid-run. TokenOps enforces one shared, run-scoped budget in real time, before the bill arrives.
An open-source desktop companion app, this weekend project questions whether working with agents should simply feel transactional.
As you add more tools to AI agents, the advantage comes from making the right tools discoverable at the right moment without paying the full cost of exposing everything upfront.
The leader of the new Microsoft Security FORGE (Frontier Offensive Research & Generative Exploitation) Labs, Kim takes us behind the scenes of codename MDASH and explains why a multi-model agentic approach may be the key to winning the security game.
With geometry, telemetry, and language feeding into a single simulation, it’s the time spent making architectural decisions that matters.
What happens when one of the founders of Writely sets up a complex agentic loop and lets it run largely independently in pursuit of an online, collaborative word processor?
Where does a judgment step go inside an app that already has a deterministic state machine, a tool surface, and a CI pipeline?
In the agentic AI era, the real durable asset isn’t the model you rent. It’s the learning loop you own. And OpenEnv is the right protocol for reinforcement learning environments (RLEs).
Developed in collaboration with GoDaddy, Google, Hugging Face, and others, this new open specification lets AI clients find resources relevant to a given task automatically.
A deterministic security system, information-flow control (IFC) offers a promising path towards secure and autonomous agents.
Long-time listener, first-time caller. Dive into a detailed conversation that breaks down the news from Build and what it means for the developer ecosystem.
Microsoft partnered with Lectures on Tap to present Build on Tap, where CTO Kevin Scott addressed a growing perceptual disconnect in the tech industry.
Agentic AI is here, and the traditional software development lifecycle is broken. Get ready for what comes next by shifting your workflows and your mindset.
New interaction technology enables new types of computers.
Omar Shahine and Jakob Werner talk shop about building an always-on, enterprise-secure, personal AI agent.
By reframing the challenge of AI agent optimization, we arrived at a new understanding of what moves the needle and built out a plausible and novel solution.
Today at Build, we introduced Web IQ, a new grounding system for the agentic web.
Adaptive Spec-driven Scoring for Evaluation and Regression Testing (ASSERT) is an open-source framework for converting natural language behavior requirements into executable evaluations of AI models and agents.
ACS is an open, vendor-neutral standard that defines how runtime governance is applied across the agent lifecycle, independent of framework, runtime, or policy engine.
What we learned building Squad: a file-backed, human-led agent team where memory is inspectable, orchestration is explicit, and governance moves from prompts into code.
Agent loops, workflows, and harnesses let developers choose the right architecture to support simple assistants and complex multi-agent systems.
By looking at Speed, Ease, Quality, and Thriving, EngThrive aims to make it faster and easier for engineers to do great work.
We sat down with Microsoft Director of Science Mukul Singh to learn how his team leveraged two years of research to bring agentic capabilities to Excel and beyond.
