Every AI agent today pays a hidden tax. When an agent needs to look something up, it visits a website, parses HTML, handles CAPTCHAs, respects rate limits, and extracts meaning from a page designed for human eyes. Seltz, a stealth startup founded by former Google Search and Perplexity engineers, just raised $12.5 million in seed funding led by Greylock to eliminate that tax entirely. Their bet: the next billion web requests won't come from browsers. They will come from agents.
The insight is brutally simple. Search engines today index the web for humans. They rank pages by relevance to a query, return blue links with snippets, and expect a person to click, read, and interpret. Agents do not work that way. An agent does not want a link to Bloomberg's homepage. It wants the current stock price of NVIDIA, verified within the last 60 seconds, in a structured JSON field it can pass to another function. It wants citation trails and provenance metadata, not advertising-supported link farms. Seltz is building the index from scratch for that programmatic consumption pattern.
What Seltz Actually Builds
Seltz's core product is a search index designed for machine consumption, not human browsing. Instead of ranking pages by SEO signals and user engagement metrics, it ranks information by freshness, structure, and verifiability. Every result comes back in clean JSON with provenance metadata: where the data originated, when it was last verified, and what confidence score the system assigns to it. The index is built from the ground up for API-first access, which means no HTML scraping, no browser automation, no workarounds.
The company operates in stealth mode, but early design partners offer some clues about the architecture. Customers include agent frameworks like LangChain and CrewAI, plus vertical AI applications in legal research and code analysis. Usage-based pricing replaces the ad-supported model that funds traditional search. This is a deliberate design choice: when agents are the consumers, the unit of value is a verified fact, not an eyeball on a page.
Why Search Needs to Be Rebuilt for Agents
The current web was designed for the browser. HTML, CSS, JavaScript, and images all optimized for a human with a screen, a mouse, and a tolerance for latency. Agents do not process those formats efficiently. Every time an agent fetches a web page, it spends tokens parsing markup, filtering noise, and extracting signal. CAPTCHA systems designed to block bots block legitimate agent traffic. Rate limits intended for scrapers throttle automated research. The entire stack penalizes the very use case that is growing fastest.
Seltz addresses this at the index layer rather than the scraping layer. Instead of building better scrapers or browser automation tools, they are building a parallel web index that serves agent-native content. This is a fundamentally different approach from tools like Firecrawl or Browserbase, which optimize the scraping pipeline. Seltz optimizes the data itself. The distinction matters because scraping tools inherit the limitations of the source material. An agent-native index, by contrast, can guarantee structure, freshness, and provenance in ways that scraping cannot.
The Market Signal for Agent Infrastructure
Greylock leading a $12.5 million seed round for a search engine with no product launch may seem aggressive until you look at the downstream economics. Enterprise AI teams currently spend an estimated 30 to 60 percent of their agent token budget on web retrieval, most of it wasted on parsing noise. If Seltz eliminates that waste, the savings compound across every agent workflow in the organization. For an enterprise running 10 million agent calls per month at an average of 2,000 tokens per retrieval, the cost difference between optimized and unoptimized web access can run into millions of dollars annually.
This is part of a broader pattern. Infrastructure built for human consumption is being retrofitted for machine consumption across every layer of the stack. Databases are adding vector support. APIs are adopting MCP standards. Authentication is shifting from OAuth flows to agent-compatible credential wallets. Search, the oldest layer of the internet, is next. Seltz is not the only player in this space, but they are the first to attack the problem at the index layer rather than the retrieval layer, and that architectural choice gives them a structural advantage.
What This Means for Founders
For founders building agent-based products, Seltz signals that the web retrieval bottleneck is about to get much cheaper. If agent-native search achieves even a 10x cost reduction in web access, entire categories of agent applications become viable that are currently uneconomical. Real-time competitive analysis across thousands of sources, automated legal research spanning multiple jurisdictions, continuous security monitoring of open-source dependencies all of these workflows currently price themselves out of viability because web retrieval is too expensive and too slow.
For founders building in adjacent infrastructure, the message is clear: every layer of the internet stack will be rebuilt for agent consumption. The companies that treat agents as first-class citizens of their architecture, rather than as a traffic source to be managed, will own the next generation of internet infrastructure. Seltz is betting $12.5 million on that proposition. The early design partners suggest the thesis is working.
Who this is for: AI founders building agent-based products who need reliable, structured web access; platform teams evaluating search infrastructure; and investors tracking the agent infrastructure stack. The search for agents is a new category, and the seed round is early enough that the category leader has not yet been decided.