Kevin Scott on the Agentic Web and the Future of Software Craft

The Shift from Scaling Laws to Capability Overhang

Kevin Scott observes that while scaling laws remain intact, the industry has moved past the phase of questioning their validity. The current challenge is a 'capability overhang'—a gap between the sophisticated reasoning models currently available and the relatively limited ways they are being deployed in products. The focus has shifted from building bigger models to building the infrastructure required for agents to be truly useful, specifically memory, tool-use, and action-taking capabilities.

Building the Agentic Web

Scott envisions an 'agentic web' that functions similarly to the internet, relying on simple, open protocols to allow agents to interact with diverse systems. He highlights the Model Context Protocol (MCP) as a critical development, serving as the agentic equivalent of HTTP. By standardizing how agents communicate with APIs and data sources, the industry can avoid the pitfalls of 'shipping the org chart' (Conway's Law) and create a more composable, interoperable ecosystem. Microsoft is actively pushing for internal systems to adopt these standard protocols to ensure their own agents remain flexible and scalable.

Security in an Open Ecosystem

Addressing concerns about the security of an open agentic model, Scott rejects the idea that one must choose between a closed, verticalized 'App Store' model and an insecure open model. He suggests that security can be achieved through open standards and AI-driven oversight. By giving agents identity and entitlement systems, users can grant granular permissions for specific tasks. Furthermore, he posits that agents themselves can act as security monitors, using their reasoning capabilities to detect anomalies or unauthorized activity in a user's digital environment.

The Craft of Software Engineering

Reflecting on his 41-year career, Scott draws parallels between software development and traditional crafts like woodworking. He argues that while new tools—like agents or CNC machines—change the process, they do not diminish the value of the craft. He encourages developers to maintain strong opinions about their work but to remain curious about new technologies. He emphasizes that the most successful startups today are not those building infrastructure, but those with a nuanced understanding of specific user problems who leverage existing tools to solve them effectively.