The AI Agent Every Company is About to Build | Vercel CEO Guillermo Rauch

Summary

Guillermo Rauch describes company AI agents as a new layer of internal software: not just chatbots that answer questions, but systems that use company knowledge and tools to complete work. He sees two major agent use cases emerging: helping people create software and improving company operations by making internal information, analytics, and workflows easier to access.

Vercel’s Internal Agent, “V”

Vercel’s internal agent, V, is available through Slack, where employees can ask questions or delegate tasks. It can route work to specialized agents—for example, a content-writing agent or DZero, which is connected to the company’s data warehouse and helps answer questions about customers and product usage. The goal is to give employees a simple interface to company intelligence while V coordinates the systems and agents behind it.

Rauch says Vercel first built the agent to address its own needs, then developed tools and infrastructure to make similar systems easier to create. These include:

• AI SDK: A way for developers to work with different models.
• AI Gateway: A way to access models from multiple providers, with the possibility of choosing different models for different tasks.
• Vercel Connect: A set of integrations, described as supporting more than 100 systems, that lets developers connect agents to business tools and events without granting blanket access.

The discussion frames a company’s customized agent as a potential competitive advantage: its skills, context, and workflows can become part of the company’s intellectual property.

Agent Design, Permissions, and Infrastructure

Rauch favors a single, familiar agent interface that acts as a router to specialized agents and tools, rather than forcing employees to choose among many separate bots. The underlying system can still use multiple agents and computers, depending on the task.

That architecture requires careful governance. Agent administrators should define:

• Which tools and company data each agent can access.
• Which employees can use particular capabilities.
• Safeguards, human approvals, and audit logs.
• Whether an agent acts directly or uses a person’s existing permissions.

Rauch emphasizes that agents may need their own computing environments to run code and use tools, but that this capability must be controlled securely. He also describes running agents efficiently: an idle agent can “sleep” rather than continuously consume computing resources.

Lessons from OpenClaw and the Agent’s “Soul”

The discussion highlights two ideas associated with OpenClaw:

• A configurable identity or “soul”: A soul.md-style file, or an agent instruction file, can define the agent’s role, principles, and context. Rauch sees this as a way for organizations to make an agent their own rather than relying only on a vendor’s default assistant.
• A computer and tools: Giving an agent access to a computer can expand what it can do, such as writing and running code. But that access creates security risks, so it needs isolation, limited permissions, and human review where appropriate.

The transcript refers to Vercel’s agent framework as “EVE” and describes an agent as a folder containing instructions, tools, and skills.

Skills, Tools, and Continuous Improvement

The conversation distinguishes an agent’s instructions and skills—how it should behave and perform a task—from its tools, which give it access to systems such as WordPress. Skills can be updated as teams learn what works. For instance, if an agent produces poor blog drafts, the organization can improve its writing skill rather than simply asking employees to repeat the same instructions.

A practical permissions example is an intern drafting a blog post through an agent. The agent might prepare the content without having direct publishing rights; the user’s existing WordPress permissions or a human approval step can control whether it goes live.

Agent quality should also be measured. Vercel gathers Slack reactions such as thumbs-up and thumbs-down, and can review negative feedback to identify failures or areas for improvement. Rauch compares agent evaluations to software tests: teams can create test cases for accuracy, task performance, and even qualities such as verbosity or tone.

Proactive Agents and Event-Driven Work

Rather than waiting for someone to prompt them, agents can run on schedules or respond to events. Examples include preparing a weekly report, monitoring social media, or notifying a team when a payment fails or an email arrives. Rauch’s model is to treat actions in connected systems—including messages in Slack—as events that can trigger agent behavior.

He predicts that creating an agent could eventually be an early step in starting a company: a continuing assistant that learns as the business grows and connects to more of its information and workflows.

Practical Starting Guide

Rauch recommends learning by building a small, useful agent:

1. Create an agent and connect it to a familiar communication channel, such as Slack or WhatsApp.
2. Pick one routine, time-consuming task that already involves a business system.
3. Write down the task’s scope and define what a good result looks like.
4. Give the agent only the tools and permissions it needs.
5. Collect feedback, improve its skills, and add evaluations as the workflow develops.

One Vercel example is turning an engineer’s product update into a customer-friendly changelog entry. The agent helps translate technical details into a clear benefit for customers, reducing the coordination needed between engineering and marketing.

Models and Performance

Rauch expects model costs to fall, models to become faster, and open-weight options to improve. He argues against depending on a single model provider so companies can choose among models and benefit from competition.

The right choice depends on the task: interactive agents should respond quickly, while background work can spend longer using more compute or multiple models to improve the result. Specific model names and comparisons in the subtitles may be inaccurate because they are auto-generated.

Main Speakers and Sources

• Riley Brown — Host and interviewer.
• Guillermo Rauch — Vercel CEO; discusses Vercel’s internal agent, agent architecture, security, development tools, and future model trends.
• Examples discussed: Vercel’s internal agent V, the DZero data-analysis agent, OpenClaw, and Vercel’s AI SDK, AI Gateway, and Vercel Connect.