Parallel’s Parag Agrawal: Building a New Web for AI Agents
In a Nutshell
Parallel is building agent-first search infrastructure that replaces human click data with agent feedback loops, enabling agents to perform thousands of searches per human action while delivering more accurate, token-efficient results. The company uses incremental index building and Shapley value attribution to solve the billion-to-billion query-to-webpage matching problem, positioning search as core infrastructure for all agent workflows. Their framework creates sustainable economics through differential pricing based on content quality and work value, with meaningful payments to content owners expected within 12-24 months as agent usage grows 10x annually.
These notes were generated by AI and may contain inaccuracies.
At Parallel, the view is that human click data is a bug and agent doing work with search should rely on agent feedback not human feedback. The company believes that models are really good at compressing information and can benefit from research that has gone into building models and apply it to search indexing and ranking.
Parag Agrawal, former Twitter CEO, founded Parallel Web Systems, which is scaling up agentic search for the agentic web. At Parallel, the company is building technology to allow agents to search and use the web. The founding bet was that agents would do it a thousand times more than humans ever have, requiring reinvention of both the technology that can power search for agents and the business models that go alongside it.
When Parag was at Twitter in leadership roles, Twitter was a post-product market fit extraordinarily scaled business where feedback loops were from hundreds of millions of customers using the product for 30+ minutes every day. The difference at Parallel is operating in a pre-product market fit company based on the premise that in a few years a new customer is going to show up on the internet.
The problem of web search is when wanting to find something and not knowing where it is on the web, going to a search engine like Google and the search engine hopefully surfaces the answer in the most convenient of locations. The search engine crawls the web by finding every URL, trying to read it, trying to organize that information in what might be called an index. This is a billion to billion matching problem - hundreds of billions of pages and hundreds of billions of queries over time that need to be matched across these two.
Three years ago when starting the company, the reasons it was difficult for others to compete included not having access to great feedback in terms of whether one search result is better than another. To collect that feedback at scale, there was a problem around human ratings and click data. It is expensive to crawl the entire web and index it, but the view at Parallel is that human click data is a bug - an agent doing work with search should rely on agent feedback, not human feedback.
The insight that convinced Parag this was a viable problem was figuring out a way to build the business which did not require spending all of the money on infrastructure upfront before being able to service a customer. The path is incrementally being able to build an increasingly larger and more sophisticated index over time as problems are solved for more and more customers.
When first launching the product, Parallel did not launch a search product first - they launched a search agent product first. The search agent could go crawl the web after a query arrived. For deep research, there is patience to the extent of a minute, with products that sometimes take 10 minutes of research. This is a lot of time to crawl a lot of pages if there is enough of a map to know what to prioritize crawling.
The search agents were competing with the alternative being outsourcing to humans to curate amazing data. Humans sitting on search engines are far easier to compete with than a search engine on day zero. By building a product that was a search agent to do real work on top of web data, they were able to incrementally build their index.
In the very early days, there were search agents being built for doing insurance underwriting workflows and claims processing workflows. People in sales were doing sales data enrichment. People in finance who would previously send data overnight to humans who would curate data that would feed into a modeling exercise started using Parallel instead. They were initially replacing where there was outsourced human work on top of web data to collect evaluations to run agents to figure out what search for agents should look like with empirical use cases instead of theoretical evaluations.
Parallel does not categorize themselves as a new lab because their output is not a model - their output is a complement to a model. What they build multiplies on top of a model to give the model better capabilities or the agent built with the model gets better and has superpowers. The goal is that whenever someone ships a better model, they have unlocked four more use cases where Parallel can be valuable.
The framing is a 100 billion to 100 billion matching problem - every query is essentially give me a thousand tokens from a trillion web pages on the web and make sure they're the right thousand tokens. Until recently, there were 3 seconds to throw compute at read time to do it. A product was shipped that now does it in 200 milliseconds. The work involves organizing information in memory across the memory hierarchy to access it fast, training models to select the right thousand tokens, and adapting pre-trained model architecture for this new problem while allocating compute to produce the best output with limited compute/latency budget.
Using Parallel search results in using under half the tokens in an agent, becoming more accurate and faster end to end. Every time half the tokens are used, if the model is context limited or memory limited, more problems can be done, or the same problems can be done cheaper or faster. It is at its core an optimization problem around quality, cost, and latency.
The interface changes when serving agents versus humans. Humans rely on keyword search. With agents, innovation has been needed on what an agent tells the search engine. The more that is known, the better it can be done. Models are less lazy than humans - fewer typos, better specified queries, longer queries, less for the search engine to guess what the agent might want.
The example given is asking for a public company's most recent financials like the headline revenue number. The authoritative filing with the SEC has that number on page 73 of a PDF. If Google cared about authority, the first result would be that PDF which takes 3.5 seconds to load instead of the second that another page does. The alternative earnings page requires grabbing to page 77 to find the answer. Humans are lazy and will not do the work. This creates an entire class of content on the web that loads fast with information above the fold that is 99.99% right.
With agents, there is not the same trade-off because agents are not clicking around and fumbling around a PDF. The approach is taking an excerpt from the most authoritative place on the web and bringing it to the agent's context window. With agents, the exact right highlighted tiny paragraph can be consistently pointed to across the web.
When a developer sends a query to the agentic search API, models run to figure out what the query is and enrich it to figure out how it will flow into the rest of the system. There are indexes which organize different subsets of the web in different ways - a big index, a fresh index, a knowledge graph, a structured index. The first layer crafts queries for each of these different systems, deciding which ones the query needs to go to and figuring out the query rewrite for each. Each has a big retrieval layer and then a ranking layer and then more ranking layers, boiling down tens or hundreds of billions of URLs or documents down into thousands, tens of thousands, down into specific excerpts and paragraphs with more and more bigger models running at each stage with different architectures pulling more features.
The billion to billion matching problem is viewed as a forever problem. The real question is at what point incremental optimization isn't worth the squeeze. There is no point where the work is done on improving this. In the super AGI build worldview, if quality can be pushed on across web search and in the model layer, it should be done because anytime something can be done 20% better than somebody else, that might give an edge.
There is no empirical data on the ground to support the view that model companies will own search infrastructure. The crawl that model companies do for pre-training is not particularly useful for building a fresh web index because model companies for training are not patient enough to go in a completionist way try to wait for really slow random JavaScript to load - the number of tokens per amount of compute is one or two orders of magnitude too bad. For Parallel, it is worth the extra effort because they are completionist.
Search infrastructure optimizing for agents is a real adjacency for all kinds of LLM inference. Model companies could build it or buy it. Parallel announced they are working with Google Cloud to be a search and grounding provider for their enterprise agent APIs. When attaching web search to inference with LLMs on GCP, the options are Google search or Parallel search, and Parallel search is product integrated with the integration optimized.
When talking to Parallel customers, there are background agents that monitor product or the web, and when something happens, they go off and take actions. A typical search agent will do somewhere between 5 to 20 searches even if it answers within a few seconds. If transitioning from using a chat GPT medium, every time a prompt is written it will do five to 10 searches, and as it is dialed up it will do hundreds and thousands of searches. A human action creates a multiplier on the number of searches that happen.
A human process monitoring credit risk across 10,000 small businesses monthly can be transformed into weekly agent-driven analysis. The multiplier becomes hundreds of thousands to a million web searches instead of the monthly human process. The agent creates a portfolio dashboard with action items that humans review.
Custom agents built in Notion pull from internal company data plus web data through Parallel APIs. A single prompt created a custom agent that performs tens to hundreds of web searches for every meeting. Each new agent built for different use cases multiplies search volume further.
Background agents should be deployed based on incremental value versus compute spend. Rationalization will occur, but deployment continues where incremental value justifies the compute cost. The same principle applies to search operations within these agents.
Initial product focus prioritized quality and cost over latency for the first couple of years. The approach was to achieve best quality search and agent products at every price point before addressing latency. Turbo was subsequently shipped as the fastest, highest quality agentic web search available.
Current adoption remains very early stage. Outside technical communities, the term "agent" remains unfamiliar. Some individuals operate at 100-1,000x search volume compared to three years ago, but these represent outliers rather than the norm. Cloudflare monitoring shows AI traffic approximately equal to human traffic in page reads.
Traditional internet economics based on human eyeballs and attention scarcity are breaking down. Content owners face uncertainty about monetization when agents rather than humans access their content. Subscription models lack data on whether agent visits convert to subscriptions.
Ads enable differential pricing where most queries lose money but high-value queries subsidize the system. This creates extraordinary margins. The same principle applies to Twitter where most users are subsidized by a minority of high-value users.
Content owners currently have limited options: either participate in head-end fixed fee contracts with labs (including training and liability components) or cut off agent access. Fixed price constructs don't scale with 50x inference growth while maintaining sustainable economics.
The proposed solution emphasizes efficient differential pricing based on both content quality and value of work being performed. The system must be scalable beyond bespoke deals and maintain incentive alignment so content owners want to participate.
Model companies pay for three primary reasons: fresh data at inference time for product display, training data, and liability protection for previously completed training. Payments represent combinations of these three factors.
Shapley values provide the framework for attribution across multiple sources contributing to single responses. The approach measures incremental value by removing individual content sources and measuring quality decline, then comparing against alternative compute costs to achieve equivalent quality.
Full Shapley value computation costs more than the agent spend itself, making direct calculation impractical. Instead, models are trained on simulation data where various content sources are withheld and quality impacts measured. This creates estimates that maintain theoretical soundness while being computationally feasible.
The framework ensures content owners participate in value creation: unique differentiated data commands higher payments, and high-value work accessing content generates premium pricing. Macro economics suggest allocating 2-10% of LLM inference spend to web data would exceed current web data business models outside walled gardens.
Current calculations indicate 12-24 months until the economics provide meaningful dollars across a wide range of web content owners as agent usage grows an order of magnitude annually.
The company was originally incorporated as Shapley Inc. because the founder was obsessed with the attribution problem. The name was changed to Parallel after six to eight months, reflecting both the parallel processing nature of the work and the emergence of a parallel web built specifically for AI agents.
Content creators now think in terms of two audiences: humans reading content and agents consuming content. This creates a dual publishing requirement where pages must be optimized for both human and agent consumption.
Earnings transcripts represent high-value content already flipped to agent-first consumption. Public company CEOs should optimize earnings communications for agent interpretation rather than just human listeners. API documentation similarly serves agent consumers as primary audience rather than human developers.
Level one involves simple agents using web search as familiar tools. Level two features sophisticated multi-agent systems with sub-agents that wake each other and coordinate complex workflows. Level three transitions the web from pull to push architecture where agents subscribe to web changes rather than making specific requests.
The future web enables agents to subscribe to changes: "call me if this happens" rather than "go find this for me." This allows continuous allocation of compute across the entire web on behalf of customers, responding to changes in satellite imagery, customer commentary, agent outputs, or human insights.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
