For 24 years, Amanda Silver has dedicated her career to supporting developers, with her recent efforts centered on creating AI tools. Following an extensive period contributing to GitHub Copilot, Silver now serves as a corporate vice president in Microsoft’s CoreAI division. There, she focuses on developing tools for deploying applications and agentic systems within enterprise environments.

Her primary responsibility involves the Foundry system on Azure, which functions as a unified AI portal for businesses. This role provides her with direct insight into how companies are implementing these technologies and where their deployments often encounter challenges.

In a recent discussion, Silver shared her perspectives on the current state of enterprise AI agents and explained why she views this field as the most significant opportunity for new companies since the advent of the public cloud.

**Your work centers on Microsoft products for external developers, including many startups not primarily focused on AI. How do you anticipate AI will affect these companies?**

This represents a watershed moment for startups, with an impact as profound as the shift to the public cloud. The cloud dramatically lowered barriers by eliminating the need for physical server space and reducing upfront capital expenditures on hardware. Everything became more affordable.

Now, agentic AI is poised to further reduce the overall cost of software operations. Many tasks involved in launching a new venture—from customer support to legal reviews—can be completed more quickly and cost-effectively using AI agents. I believe this will lead to a surge in new startups and ventures. We will likely see companies achieving higher valuations with smaller core teams, which is an exciting prospect.

**What does this look like in practical terms?**

We are observing multi-step agents being widely adopted for various coding tasks. For instance, developers must regularly update the libraries their code depends on, such as older versions of the .NET runtime or Java SDK. Agentic systems can analyze an entire codebase and perform these updates far more efficiently, potentially reducing the required time by 70 to 80 percent. This requires a deployed, multi-step agent to accomplish.

Live-site operations present another key application. Maintaining a website or service often means having someone on-call to respond to overnight incidents. While teams still maintain 24/7 coverage, we have developed agentic systems that can successfully diagnose and frequently resolve issues autonomously. This reduces the need for engineers to be awakened for minor problems and significantly shortens the average incident resolution time.

**Agentic deployments haven’t progressed as rapidly as some anticipated six months ago. What do you think is holding them back?**

For those building agents, a major hurdle is often a lack of clarity regarding the agent’s purpose. There’s a necessary cultural shift in how these systems are conceived. Builders must ask: What specific business problem am I solving? What is the goal? You need a clear definition of success for the agent and must consider what data you provide to enable it to reason through the task. These foundational questions are typically bigger obstacles than general apprehensions about deployment. Anyone who examines these systems can recognize their potential return on investment.

**You mentioned general uncertainty, which from the outside seems like a significant barrier. Why is it less of an issue in practice?**

Firstly, it will be very common for agentic systems to incorporate human-in-the-loop scenarios. Consider a package return process. Traditionally, this might be 90% automated, with a human inspecting the package to assess damage before approving the return. Today, computer vision models have advanced to the point where much of that human oversight is no longer necessary, though borderline cases may still require escalation—similar to calling in a manager.

Certain critical operations will always demand some level of human oversight, such as incurring a legal contractual obligation or deploying code that could impact system reliability. Yet, even in these areas, the question remains: How much of the surrounding process can we successfully automate?