Software for Agents
The transcript argues that AI agents represent the next trillion users of the internet and require purpose-built software infrastructure. Unlike humans, agents need machine-readable interfaces such as APIs, MCPs, and CLIs rather than visual UIs. The biggest startup opportunity lies not in building agents, but in building the software agents depend on.
Summary
The transcript opens with the provocative claim that the next trillion internet users will not be humans but AI agents, framing this as a foundational shift in how software must be designed. The speaker argues that agents are already performing real-world tasks — browsing the web, conducting research, making purchases, and managing CRM systems — but are forced to do so using software built for human interaction, which is described as slow, inconsistent, and brittle.
The core argument is that agents require an entirely different software foundation. Rather than visual interfaces like forms, buttons, and dashboards, agents need machine-readable interfaces such as APIs, MCPs (likely Model Context Protocols), and CLIs. Additionally, agents require comprehensive documentation that allows them to discover, sign up for, and begin using new tools programmatically, without any human involvement in the loop.
The speaker then makes a market-structure argument: every major software category used by people today needs to be rebuilt with agents as first-class citizens. Critically, this rebuilding will not come from incumbents retrofitting agent support onto existing products, but from startups that design explicitly for agents from the ground up. The transcript concludes with an investment or partnership pitch, suggesting that building software for agents — rather than building agents themselves — represents the largest untapped opportunity in the current AI landscape.
About this episode
The next trillion users on the internet won't be people. They'll be AI agents, and they're already doing real work on top of software that was designed for humans clicking buttons. Every major category of software needs to be rebuilt for agents as first-class citizens, and that won't come from incumbents. Apply to YC Summer 2026 at ycombinator.com/apply.
Key Insights
- The speaker argues that agents are already performing real-world software tasks but are doing so on top of human-designed interfaces, making their operation slow, inconsistent, and brittle — implying the mismatch between agent capabilities and current software is the core problem to solve.
- The speaker claims that agents specifically require thorough documentation enabling them to discover, sign up for, and instantly use new tools programmatically without any human in the loop — framing autonomous onboarding as a distinct and unmet infrastructure requirement.
- The speaker asserts that agent-first software will not come from incumbents bolting on agent support, but from startups that build for agents as first-class citizens — positioning this as a greenfield startup opportunity rather than an enterprise evolution.
Topics
Transcript
[0:00] The next trillion users on the internet won't be people. They'll be AI agents. And now is the time to make something agents want. Agents are already browsing the web, doing research, making purchases, and managing legacy CRM, but they're doing it on top of software that was designed for humans clicking buttons in a browser, which is slow, inconsistent, and brittle. Agents need a completely different foundation. Instead of visual interfaces like forms, buttons, and dashboards, they need [0:32] machine readable interfaces like APIs, MCPs, and CLIs. Agents also need thorough documentation to enable them to discover, sign up for, and instantly start using new tools programmatically without needing a human in the loop. That means every major…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator
Patrick Collison: "What If You Succeed?"
Patrick Collison discusses Stripe's founding story, the importance of concrete customer problems, and shares data suggesting it's the best time ever to start a business. He argues that despite AI advancement, there will be many winners rather than centralization, and emphasizes the value of deep learning alongside AI tools.
Multi-GPU Kernels, Intelligence per Watt, Heterogeneous Inference, and More | YC Paper Club
This YC Paper Club event focused on chip and kernel specialization in AI systems, featuring discussions on multi-GPU kernel optimization, energy-efficient inference, AI-generated kernel code, heterogeneous inference hardware, and GPU-accelerated game engine simulation. Speakers explored how specialization across hardware, software, and algorithms can dramatically improve efficiency and performance across different AI workloads.
Blake Scholl: The Future Was Supposed to Be Faster
Blake Scholl, founder of Boom Aerospace, discusses building the first independently developed supersonic jet by applying software development principles to hardware manufacturing. He argues that startups can compete in regulated industries by building solutions that solve regulatory concerns, and emphasizes that passion, iteration, and vertical integration are key to breakthrough innovation in deep tech.
Boris Cherny: Building Claude Code
Boris Cherny, creator of Claude Code, discusses how Anthropic's Opus 5 model represents a major leap in AI capabilities, requiring a fundamental rethinking of how to build AI products. Rather than accumulating complexity, successful builders must regularly delete system prompts and code to "unhobble" the model, allowing it to tackle increasingly complex tasks like rewriting entire codebases and maintaining software automatically.
Self-Maintaining APIs
The speaker argues that API communication is fundamentally broken across the industry, with breaking changes and new features going unnoticed. Since agentic coding tools have proven viable and developers accept external tool access to codebases, API providers should automatically apply changes to customer code rather than just announcing them—similar to how Dependabot works for dependencies.