Kevin Scott on The Future of Programming, AI Agents, and Microsoft’s Big Bet on the Agentic Web

The Agentic Web: Infrastructure and Protocols

• 💡 The agentic web requires agents to take action, use tools, and access diverse information, necessitating an ecosystem similar to the internet.
• 🔑 Microsoft is standardizing internal systems with Model Context Protocol (MCP), which acts as the "HTTP of agents" for seamless tool and data access.
• 🌐 NLWeb is Microsoft's initiative to make websites legible to agents, akin to how HTML made content accessible for browsers.

Security and Open Innovation

• ✅ The notion of a trade-off between permissionless innovation in open systems and robust security is a false dichotomy.
• 🛡️ AI agents can provide enhanced security by understanding user preferences and communication patterns to detect and act on suspicious activities.
• 💬 Agents could triangulate information across communication modalities to verify legitimacy of activities, as exemplified by a personal anecdote.

Craftsmanship in the Age of AI Agents

• 🛠️ Makers, including software engineers, often have strong opinions about their craft, tools, and processes.
• 🧠 Kevin Scott, with 41 years of programming experience, advises curiosity and experimentation with new tools, using them if they work and discarding them if they don't.
• 🎯 The choice between valuing the process versus the outcome often dictates tool adoption, with some prioritizing efficiency and others the traditional craft.

The Future of Software Engineering with Agents

• 🚀 The future will feature a diverse ecosystem of agents, not a single dominant tool, allowing developers choice and flexibility.
• 💡 Problem-solving will drive agent diversity, with companies building specialized agents that deeply understand and address specific user needs.
• 📈 Agentic workflows will evolve from synchronous interactions to longer, asynchronous feedback loops, enabling agents to tackle more ambitious tasks.

Addressing Agent Limitations

• 🧠 There's a "capability overhang" where models' reasoning abilities exceed their current product applications, requiring industry effort to bridge the gap.
• ⏳ Current agents suffer from constrained memory, making them transactional and limiting their ability to handle increasingly complex, delegated tasks effectively.