Amanda Silver of Microsoft has dedicated 24 years to empowering developers, a focus that has recently shifted to creating AI tools. After significant contributions to GitHub Copilot, Silver now serves as a corporate vice president within Microsoft’s CoreAI division. In this role, she concentrates on developing solutions for enterprise application deployment and agentic systems. Her efforts center on the Foundry system within Azure, designed as a unified AI portal for businesses. This position offers her unique insights into how companies are actually utilizing these systems and where deployment challenges often arise.
I had a discussion with Silver regarding the current capabilities of enterprise agents and why she views this moment as the most significant opportunity for startups since the advent of the public cloud.
This conversation has been edited for brevity and clarity.
Your role at Microsoft involves creating products for external developers, many of whom are startups not primarily focused on AI. How do you envision AI influencing these companies?
I believe this is a pivotal moment for startups, as impactful as the shift to the public cloud. Consider the immense effect the cloud had: it freed startups from needing physical space for server racks and reduced capital expenditure on hardware for their labs. Everything became more affordable. Now, agentic AI will further decrease the overall costs associated with software operations. Many tasks involved in launching a new business—from support to legal inquiries—can be performed more quickly and economically using AI agents. I anticipate this will lead to the launch of more ventures and startups. Furthermore, we’ll likely see higher-valuation startups operating with smaller teams. I find this prospect quite exciting.
Could you provide practical examples of this?
We are certainly observing the widespread adoption of multi-step agents across various coding tasks. For instance, developers must constantly update their codebase with the latest versions of dependent libraries. If you rely on an older version of the .NET runtime or Java SDK, agentic systems can now analyze your entire codebase and update it much more easily, potentially reducing the required time by 70% or 80%. This truly necessitates a deployed multi-step agent to accomplish.
Live-site operations are another example. Imagine maintaining a website or service where an issue arises unexpectedly, requiring an on-call person to be awakened to address the incident. We still have personnel on call 24/7 in case a service goes down. However, this used to be a highly unpopular duty because minor incidents would frequently disrupt sleep. We’ve now developed an agentic system that successfully diagnoses and, in many cases, fully resolves issues that emerge in these live-site operations. This means humans no longer need to be disturbed in the middle of the night to groggily diagnose problems. This also significantly reduces the average resolution time for incidents.
One current enigma is that agentic deployments haven’t progressed as rapidly as anticipated, even just six months ago. What do you believe is the reason for this?
If you consider the individuals developing agents, a common barrier to their success is often a lack of clarity regarding the agent’s intended purpose. There’s a necessary cultural shift in how these systems are built. What specific business problem are they trying to solve? What outcome are they striving for? You need to have a very precise definition of what constitutes success for this agent. Additionally, you must consider what data you are providing the agent so it can effectively reason and accomplish its assigned task.
We perceive these issues as more significant obstacles than the general apprehension surrounding agent deployment. Anyone who examines these systems quickly recognizes the return on investment.
You mentioned general uncertainty, which from an external perspective seems like a major impediment. Why do you view it as less of an issue in practice?
Firstly, I believe that human-in-the-loop scenarios will be very common for agentic systems. Consider a package return process. Previously, the workflow for processing returns was 90% automated and 10% human intervention, where someone had to inspect the package and make a judgment call on the extent of damage before approving the return.
This is a perfect example where, now, computer vision models have become so advanced that in many cases, we don’t need as much human oversight for package inspection and determination. There will still be some borderline cases where computer vision might not yet be sufficient to make a decision, potentially requiring escalation. It’s akin to how often a manager needs to be called in.
Certain operations will always necessitate some form of human oversight due to their critical nature. Think about incurring a contractual legal obligation or deploying code into a production environment that could impact system reliability. However, even in these instances, the question remains: how much of the process can still be automated?