Building the Agent-Native Company: A Conversation with Vercel's Guillermo Rauch

Introduction

This interview features Guillermo Rauch, CEO of Vercel, in conversation with Riley, exploring how AI agents are transforming business operations. The core theme centers on the emergence of agent-native organizations, where companies deploy internal AI agents like Vercel's "V" to democratize knowledge, automate workflow, and reshape the fundamental architecture of how businesses operate.

Structured Summary

🧠 The "Brain Agent" Concept

Vercel's internal agent, named "V," serves as a company-wide intelligence layer that approximately 1,000 employees interact with daily through Slack.

• Universal accessibility: Any employee can simply type "@V" in Slack to access company knowledge, customer data, and analytical capabilities
• Sub-agent architecture: V coordinates specialized sub-agents, including a content agent for marketing and "DZero" for data analysis connected to their data warehouse
• Beyond support: Initially conceived as a support assistant, V evolved into a fully operational agent capable of delegating tasks to coding tools like Codex or creating prototypes with v0
• The "killer app" insight: Rauch identifies two primary agent use cases: software development and what he calls the "run your company better" agent—a knowledge-base-plus-analytics system that helps employees navigate organizational information and business intelligence

🏗️ The EVE Framework

Vercel developed EVE as a framework to democratize agent creation, inspired by lessons from open-source projects like OpenClaw.

• The "soul" concept: An EVE agent fundamentally consists of a folder with an instructions.md file—markdown text defining the agent's genesis, principles, and purpose
• Simplicity by design: Rauch emphasizes that agents don't need to be complex; they can be as straightforward as a folder hierarchy with tools, skills, and instructions
• Learnings from OpenClaw: Key insights included the power of coding agents, the importance of giving agents dedicated computers, and crucially, the value of customizing agents beyond off-the-shelf models like Claude
• Serverless agent operation: EVE agents can hibernate when inactive, similar to a Mac Mini sleeping when not in use, making large-scale deployment economically viable

⚖️ God Agent vs. Team of Agents

Rauch advocates for what he describes as a "god model" approach, drawing parallels between AI architecture and how corporations provision devices.

• The user experience imperative: The ideal is Star Trek's computer or Iron Man's Jarvis—ambient computing where users don't target specific capabilities
• Router architecture: V acts as an intelligent router, directing queries to appropriate capabilities based on context—knowledge base, support tickets, or data analytics
• Governance controls: The new "IT department" job is defining what capabilities to bundle into the agent and managing identity, permissions, and access control
• Practical metaphor: Just as a corporate phone comes preconfigured with applications, an internal agent should come preconfigured with organizational capabilities, with access levels determining what information different employees can access

🔧 Practical Implementation

For companies wanting to build their own agents, Rauch offers concrete guidance based on Vercel's experience.

• Starting point: Connect an EVE agent to a primary communication channel (Slack, WhatsApp, or iMessage) and choose one "toil task" that has a clear system but would benefit from automation
• Example use case: Vercel automated their product changelog process—engineers create a Slack thread, and the agent refines their technical input into clear customer-facing language using a formula: "what's the benefit, how much does it cost, and what do I do to get it"
• Iterative improvement: When feedback indicates poor output (e.g., "Claude slop"), the response isn't reprimanding employees but improving the agent's skills through updated instructions files
• Feedback loops: Every agent response includes thumbs up/down UI, with nightly jobs aggregating negative feedback to propose self-improvements
• Evaluation system: EVE supports evals—test cases that assess logical soundness, information accuracy, and even personality traits (like being too verbose)

🚀 Future Outlook

Rauch identifies several emerging trends that will accelerate agent adoption and capability.

• Falling cost of intelligence: He highlights dramatic price reductions, mentioning GLM models and Kimi K3 as examples of increasingly capable open-weight models at lower costs
• Model diversity and competition: Vercel remains model-agnostic, allowing customers to benefit from competition across providers while retaining ownership of data and skills
• Speed improvements: Fast models will become dramatically faster, with Rauch noting that "GLM 5.2 to fast is astonishingly fast and it's only getting faster"
• Proactive, event-driven agents: The future includes agents that work autonomously in the background—reading social media, parsing keywords, generating reports, and providing daily briefings without direct prompting
• Batch intelligence market: Vercel is building capabilities for asynchronous inference where agents can submit tasks without caring about completion time, creating something like a "spot market for intelligence"

Concluding Insight

The fundamental shift is from prompting agents to owning them—creating "an intelligence of your own" rather than relying on third-party models. As Rauch notes, just as entrepreneurs register domain names as their "fingerprint" on the internet, the future company will build its agent before building its website, making agent ownership as fundamental to business identity as incorporation itself.