Reworkd pivoted from viral AgentGPT to AI web scrapers and raised $4M total
After AgentGPT hit 100,000 daily users in a week and cost $2,000 a day in API calls, Reworkd narrowed to AI agents that scrape structured data.
What was changed
In early 2023, Asim Shrestha, Adam Watkins and Srijan Subedi were still living in Canada when their free browser tool AgentGPT — a simple interface for building autonomous AI agents — went viral on GitHub, acquiring more than 100,000 daily users within a week. The influx caught them off guard: COO Subedi said the tool was costing $2,000 a day in API calls, so they had to incorporate as Reworkd and get funded fast, earning a spot in Y Combinator's summer 2023 cohort.
The founders quickly concluded that building general AI agents was too broad. But the most popular use case on AgentGPT was creating web scrapers — a relatively simple but high-volume task — so Reworkd made that its singular focus: AI agents that extract structured data from the public web. Customers hand over hundreds or thousands of URLs plus a description of the data they want; the agents use multimodal code generation to write unique scraping code per site, in a 'self-healing' approach meant to survive page updates.
By July 2024 the four-person company had raised $2.75 million in seed funding from Paul Graham, AI Grant (Nat Friedman and Daniel Gross's accelerator), SV Angel, General Catalyst and Panache Ventures, on top of a $1.25 million pre-seed, for $4 million total. Its marquee customer Axis uses the agents to extract data from thousands of government regulation documents across the EU; the team avoids scraping news and sign-in walls entirely to steer clear of the copyright fights dogging OpenAI and Perplexity.
Why it worked
Virality became a liability: 100,000 daily users cost $2,000 a day with nothing monetized, forcing the founders to incorporate and fundraise in a hurry.
General agents were too broad a bet for a tiny team, but scraper generation was already the most popular AgentGPT use case — demand was proven inside their own product.
The AI data hunger gave the pivot a market: building AI models was the number one use of public web data in 2024, and human-built scrapers for thousands of small sites were uneconomic.
Riding model progress instead of competing with it — GPT-4o's multimodal code generation made the product possible only months before — kept a four-person team viable.
What can be applied
A viral product is not a business: when usage outpaces your ability to serve it, pivot to the one use case inside it that customers already need.
Aftermath
The company stays deliberately tiny — four people, with founding research engineer Rohan Pandey described by an investor as 'a one person research lab' — while carrying heavy inference costs it expects to fall as smaller models like GPT-4o mini improve. It built an open-source evaluation framework to catch wrong scrapes, avoids news aggregation by choice, and leans on court rulings like the Bright Data verdict holding public profiles fair game. Paul Graham and AI Grant did not respond to TechCrunch's request for comment.