One God Agent vs a Team of Agents

Riley: Really important thing about OpenClaw is soul.md — the soul of your agent that’s going to help you run your company. What I believe will happen: even before you build a website, you’re going to build the agent that helps you build the company. Most of the world still thinks about agents as something you prompt. Can we automate even the prompting so the agent does useful work while I’m not at the computer?

Today I’m talking with Guillermo Rauch, CEO of multi-billion-dollar company Vercel, about how companies use agents — Vercel’s internal agent used by almost a thousand people; one god agent vs a team of many; challenges of setup; open models like Kimi; and how operators and individuals use agents to be more productive. Agent Native. Let’s dive in.

Guillermo, thank you. My audience is mostly business operators. With models from China and the US, Claude, Fable, Opus-class — where are companies at implementing agents?

Guillermo: Think of the agentic revolution like any platform shift — look at killer apps. PC: word processor (and games). Mobile: not only smaller screens but new use cases. With agents: number one killer app is building software — coding agents. Coding is a core capability for knowledge work: data science for a deck, automation, summarization into slides.

Levels of expertise: agentic engineering (I’ve programmed 20 years; I use a coding agent on hard tasks); vibe coding (prototypes and full-stack via v0, Lovable, etc.). Another killer app: the “run your company better” agent — knowledge base + data analysis + project management — a brain that navigates org charts and overwhelming internal systems. At Vercel, anyone helping a customer or closing a sale or building software needs: when did they reach out, time spent, platform usage, how many SQ of Vercel. That internal brain agent is a killer app of AI.

Riley: I’ve tried this with a nine-person marketing team. Personally I’m in Codex for knowledge work. How do I enable this for a team? New hires need skills, but many skills hit personal email / permissions. What’s the interface at team scale?

Guillermo: Codex and ChatGPT give a taste. The new job of someone who runs a company is to enable the workforce with agents and work on the agent. Future IP / competitive edge is create, tune, optimize, and disseminate agents internally.

Our internal agent is V. Anyone in Slack says @V for day-to-day. Marketing promoting a product, product change, eng blog with an engineer — through V. V has skills we update continuously, sub-agents, a virtual employee team: content agent; data analysis agent we call D0 — nexus of intelligence, connected to the data warehouse. Join Vercel, join chat, omnipresent intelligence. V can coordinate: delegate to Codex, prototype with v0, query production. Key is every company deploying this brain and optimizing it over time.

Riley: Is there a team that manages V? How big? Day to day?

Guillermo: Product philosophy: solve it ourselves first. Inspiration: anytime you talk to someone, could an agentic layer have gotten you the info sooner? ChatGPT reasons but lacks internal KB, customer records, how we build software.

We built AI SDK (any model), AI Gateway (tokens from any model). For V we don’t want it locked to Claude or Codex — autonomously choose the best model per task. V as a superset of agents. Designated a few people for the conversational experience. First: support assistant — hiring, Slack with customers, @Vercel-style knowledge of how Vercel works. Difference: assistant vs agent — agent does things. Skills and jobs to be done. Named V for internal; customer-facing distinct as @Vercel. Created the V team.

Hard part: assemble tools, frameworks, infrastructure and improve over time. We built V, then shared the framework: Eve. Next.js/React made web apps easy; every knowledge worker will want an agent of their own — Eve enables that.

Riley: OpenClaw feels fad-like on Trends, but unlocked a magic moment — agent on a computer doing things.

Guillermo: OpenClaw informed Eve. Fundamentally: raw model intelligence plus every tool — write code, run it, access everything. Magic to the point of accidents (Peter’s story: found an API key and did something unasked). Heartbeat: wake without a prompt. soul.md: not just Anthropic’s agent — yours. Eve agent at basic: folder with instructions.md — soul of the agent that runs your company. Computer matters (Mac Mini era); also dangerous — needs space and security. Hire a knowledge worker: IT gives a laptop with apps and logins. Same for agents, but secure: define tools, human-in-the-loop, data access. Not every agent needs 24/7 computer — serverless sleep when idle (hibernating Mac Mini metaphor). At our scale we need efficient agent ops.

Riley: Labs struggle to convey “agent with a computer” vs chatbot; products differ by computer access (chat vs Work vs Codex terminal).

Guillermo: Whether intern or senior eng, computer count is implementation detail for the builder. Governance and data access is the new job — agents touch customer data. Users talk to orchestrating agent that may use a million computers or none. Outcomes and accuracy matter; performance / delight matter more and more.

Riley: Creating a V for my team — god agent that knows everything vs team of agents sharing a knowledge base? Marketing vs finance permissions.

Guillermo: I’m a UX guy. Ideal UX is Star Trek computer / Iron Man Jarvis — ambient, don’t target a specific capability. Hundreds of internal tools exist that I don’t know. Intelligent agents act as routers. V is a router: docs/knowledge path; support agent with ticket infra; etc. So more on the god-model experience. Metaphor: corporate phone preconfigured with identity and apps. New IT: which capabilities into the agent, and identity/access — business-specific and regulation-specific. Early startup: almost equal access; early Vercel maybe two people had financials, most of first ten had read access to almost everything. Foundational agent owner sets access control, tools, guardrails, audit trails — hard work; framework makes that the job so model/infra is offloaded.

Riley: Agent sees who/where the message is from and scopes accordingly.

Guillermo: Most agents are built by talking to an AI that helps maintain the Eve project. Think hierarchy of files/folders. Start with instructions: you help run Riley’s business, values… Tabula rasa + raw model intelligence. Marketing content / blog: create tools folder, expose WordPress read/write, ship. Channel: WhatsApp, Telegram, Slack, Teams, iMessage. Intern shipping to prod? Probably not — agent admin sets org-based rules, or WordPress auth stays with human and agent only drafts. Unblock drafts → then “this is Claude slop” escalation → you improve the content-writing skill, not only scold the intern. Meta-work: put intelligence into skills and tools. Ground with social (plugins like ScrapeCreators). skills/contentwriting.md: what we like / don’t. Schedules: nightly social → drafts or Slack report. Proactive info: every Monday download across product areas. Most people still only prompt; alpha is work while offline.

Riley: Team-level external triggers?

Guillermo: Events from Stripe (refund, failed payment), email, etc. Hard part isn’t the POC — it’s securely connecting systems. Vercel Connect: 100+ systems with developer control, not full access day one. Subscribe to events; Slack is just another event. World as events.

When I started Vercel, earliest fingerprint was a website. Future: before the website, you build the agent that helps build the company — factory, trusted advisor, learning the trajectory. Self-improvement: evals, tests/exams so sophistication progresses. More fundamental than the .com. Slack as agent OS; thumbs up/down; nightly job aggregates negatives and proposes self-improvement — humans still in the loop; framework evals for accuracy and personality (e.g. too verbose).

Riley: Next few years — only technical people building company agents?

Guillermo: First building block everyone can create (like drag-drop websites). Business-critical agents: everyone contributes skills, critique, feedback; engineers on core loop, data access, governance, security. Systems design / threat models / data flows more than pure code writing.

Riley: Tangible start for non-technical owners?

Guillermo: Create Eve agent at eve.dev, connect favorite chat medium, pick one toil task with a system. Example: product changelog — used to need eng × marketing coordination; now one Slack thread, agent refines to context-free customer language (benefit, cost, what to do — seven seconds shave story). Put your secret sauce into the skill. Ship on Vercel.

Riley: Next 3–6 months unlock?

Guillermo: Cost of intelligence down; variety of models; model-agnostic access so you keep ownership of data/skills and benefit from competition. Models get way faster — today’s AI is first-iPhone slow. Interactive: want fast. Async: want accuracy; batch tokens / spot market for intelligence (SETI@home-like spare capacity). Open weight competition extreme — GLM 5.2 strong, fast variants; Kimi raises the bar, cyber evals strong. Workhorse interactive: Grok / GLM price-performance. Night exec insights: more reasoning; multi-model consortium.

Riley: Business owners resist Claude Tag-style lock-in.

Guillermo: Beyond the model — own an intelligence of your own. Baptize your agents: this is our agent, our company. Like owning your domain on the web — intelligence age version of that.

Thanks — anytime, Riley.