Overview of 20VC with Parag Agrawal, Founder of Parallel
Harry Stebbings sits down with Parag Agrawal, former Twitter CEO and founder of Parallel, to discuss how agentic search will reshape the web, why traditional ads-based monetization breaks in an agent world, and what the next few years of AI infrastructure may look like. The conversation covers search architecture for agents, model scaling, data markets, vertical integration, security/guardrails, and Parag’s lessons from working with Elon Musk at Twitter.
The Core Idea Behind Parallel
“Google for agents”
Parag describes Parallel as “the Google for agents” — infrastructure that helps AI agents search the web efficiently.
- Agents will use the web far more than humans do, potentially by 1,000x.
- That changes both:
- the technology stack required for search
- the business model used to monetize it
- The company is building for an agent-native future where:
- inputs are longer and more explicit
- outputs are not blue links, but tokens, files, or structured actions
- latency and compute budgets vary dramatically by use case
Why Search for Agents Must Be Different
Humans and agents have different needs
Parag argues that agent search cannot simply be “human search with a new interface.”
- Humans search with short, underspecified keywords and tolerate about 0.5–1 second latency.
- Agents may:
- ask in full sentences
- need answers in 100ms for voice workflows
- or be background workers that can think for seconds and optimize for quality
- Because agents use search so much more, search must be 10x–100x more efficient than current systems.
Compute allocation is the key problem
The core technical challenge is allocating compute efficiently:
- Start with billions/trillions of web documents
- Narrow down to a small set of relevant sources
- Spend just enough compute at each stage to produce the best answer at the lowest cost
- Different models require different tradeoffs:
- cheap models can tolerate more imperfect search
- expensive frontier models justify heavier search investment
Model Scaling and the Future of AI
Frontier models will keep getting bigger
Parag believes frontier models will continue to scale upward.
- He sees no clear end to the scaling law
- Frontier models will likely keep becoming more capable as more compute is thrown at harder problems
- At the same time, smaller models will keep reaching the performance level of today’s larger models
The market will fragment by use case
He expects a broad spread in model sizes and capabilities:
- frontier models for the most valuable tasks
- smaller, cheaper models for many common workloads
- model choice will depend on the task, latency tolerance, and cost sensitivity
Why Ads Don’t Work in an Agent World
The ads model breaks down
One of Parag’s strongest opinions is that ads are not viable in their current form for agents.
- If an agent reads content on a publisher site, the human may never see the ad
- In e-commerce, agents may bypass the traditional ad-driven buying journey
- If agents become the primary customer, ad inventory loses value
A replacement is needed
Parallel is building something like “AdSense for agents”:
- pay content owners when agents derive value from their information
- compensation should be based on marginal contribution
- this is meant to incentivize content providers to allow agent access rather than block it
Data pricing is a major missing market
Parag argues the industry still does not know how to properly pay for:
- unique data
- proprietary insight
- inference-time access to valuable sources
He sees this as a huge future market, especially as intelligence becomes cheaper and more value comes from combining:
- proprietary data
- public data
- agent reasoning
The Future of Content, Publishers, and Data Access
Every company must decide whether to allow agents in
Parag believes most businesses will eventually have to let agents access their systems, but on negotiated terms.
Examples mentioned:
- Amazon blocking some agent access
- Shopify and Expedia being more open
His view:
- companies with market power may try to restrict agents
- but over time, agent access becomes inevitable
- the key question is who gets paid and how
Publishers need new economics
For content businesses like news, the current setup is unsustainable if agents consume content without monetization.
- Agents should not read content for free if they create value
- content owners need a direct compensation mechanism
- otherwise, they will block bots and agents
Margin, Revenue, and Scale
Parallel sees a massive market opportunity
Parag argues Parallel sits adjacent to the broader AI inference market.
- a meaningful share of inference spend will go to web search
- he estimates 5%–20% of inference GPU spend may flow into search infrastructure
- as inference spend grows, Parallel can grow with it
The business can scale very quickly
He believes the company can become extremely large if:
- agent adoption continues
- Parallel maintains technical leadership
- partnerships with content owners scale
- customers trust it for mission-critical workflows
He suggests a path to very large revenue in a few years is plausible if execution remains strong.
Pricing Will Likely Fall Dramatically
Web search is overpriced for agent use cases
Parag thinks current pricing in web search is distorted by the human-search era.
- If a cheap model uses a lot of web search, search can become the majority of the cost
- That is backwards: search should usually cost less than the model
- He believes search can become 10x cheaper again while maintaining quality
Better tech, lower cost, more usage
He wants a race to the bottom on price if it comes from better technology.
- cheaper search enables more agentic use cases
- lower cost creates more demand
- this is a Jevons Paradox-style expansion dynamic
The Future of Search: From Pull to Push
Search will become event-driven
Parag’s view of the future is not just “agents query search.”
Instead, he expects:
- always-on agents
- persistent monitoring
- web crawling as a continuous process
- alerts and notifications triggered by changes in the world
Example:
- instead of polling the web every six hours for a young founder
- the system can continuously monitor changes
- then trigger only when relevant events occur
This shifts web search from pull to push.
Risks, Guardrails, and Security
Agent capability creates security risk
Parag acknowledges the danger of powerful agents, especially in cybersecurity.
- models can be used for hacking and abuse
- some incidents happen because guardrails were insufficient
- alignment remains unsolved, but current efforts do reduce risk
The real problem is misuse, not just capability
He distinguishes between:
- accidental misuse due to poor safeguards
- deliberate misuse by malicious actors
His concern is that society may:
- adopt AI too quickly
- fail to diffuse the benefits broadly
- create serious disruption before adapting
Personal Reflections and Lessons from Elon Musk / Twitter
What Parag learned from Elon Musk
He says the most striking thing about Elon is:
- urgency
- the ability to compress time
- pushing people to do more than they think they can
He sees that as a real founder skill, even though he has disagreements with Elon.
Biggest lesson from Vinod Khosla
From investor Vinod Khosla, Parag learned:
- keep a strong technical intuition
- think long term
- preempt the next two or three technical bets after solving one problem
What he changed his mind on
In the last year, Parag says he underestimated:
- the value of strong sales
- the importance of marketing
He used to focus almost entirely on product and technology, but now sees those go-to-market skills as materially important week to week.
Final Takeaways
Parag’s big beliefs
- Agents will use the web dramatically more than humans
- Search for agents needs a new technical architecture
- Ads will not survive in their current form in an agent-native world
- Data and insight will need new pricing and transaction models
- Search will move from pull to push
- Frontier models will keep getting bigger, while smaller models will remain extremely useful
- Security, alignment, and social adaptation lag behind technical capability
The main message
The internet is moving from a human-first system to an agent-first system. Parag’s thesis is that whoever builds the best infrastructure for agentic search — and the right economics around it — will be positioned at the center of that transition.