20VC的哈里·斯特宾斯邀请Parallel的创始人帕拉格·阿格拉瓦尔，就代理搜索、人工智能的未来和不断演变的商业模式进行讨论，该内容发布于2026年9月26日。 帕拉格将Parallel介绍为“代理的谷歌”，它提供专为AI代理设计的网络搜索能力。他强调，代理使用网络的频率将是人类的1,000倍，这需要全新的技术和商业模式。与输入简短查询并期望10个蓝色链接的人类不同，代理需要不同的输入（完整句子）、不同的时间限制（语音代理需要100毫秒，而后台任务没有时间限制）以及不同的输出（令牌或文件）。Parallel优化了网络搜索中的计算资源分配，以节省模型上的计算，为廉价的Luna或昂贵的Fable等模型定制响应。该公司的API允许代理指定其需求（例如，高精度、低延迟），提供“Turbo”模式用于提速和“Advanced”模式用于深度、昂贵的查询。 Parallel的主要用例涵盖广泛的知识工作，包括编程、AI律师、保险承保人和科学家。尽管编程使用较少的网络搜索，但法律和销售等领域高度依赖网络搜索。帕拉格认为，随着模型变得更好、更便宜，从而实现更多的代理用例，Parallel的业务将蓬勃发展，因为他们押注代理将成为网络的未来消费者。 在讨论模型趋势时，帕拉格预测前沿模型将继续变得更大、能力更强，而小型模型将更高效地达到固定的性能水平。他挑战了90%的令牌活动将通过开放模型进行、90%的资金将流向前沿模型的观点，强调了美国开放模型和竞争的价值。他认为，由于GPU供需不平衡，模型路由层目前具有价值。 展望未来3-5年，帕拉格认为数据和独特的见解极具价值，并强调了将它们货币化用于代理的挑战（例如，风险投资家使用的代理对PitchBook数据的需求）。他认为像亚马逊这样最初会抵制代理的公司，最终不得不允许它们进入，但参与条款将至关重要。 帕拉格的一个重大担忧是，随着代理成为主要客户，广告行业将瓦解。由于代理不看广告，他提出了“代理的AdSense”概念，即Parallel根据代理从其内容中获得的利益向内容所有者支付可变金额，从而调整激励措施，并将网络消费从“拉动”转变为“推动”。他认为，这将为内容创造一个更大的整体市场。 Parallel作为一家基础设施企业运营，致力于提供卓越的质量、更低的成本和更快的延迟。帕拉格断言，当前的网络搜索定价与代理需求不符，往往比必要的价格高出50倍。他的目标是让Parallel以当前市场价格的1/10提供网络搜索，并有可能再降低10倍，他认为价格上的“触底竞争”对于实现1,000倍的规模至关重要。他解释说，Parallel的技术能以传统以人为中心的网络搜索计算成本的1/20到1/50，实现相同的质量。帕拉格预测Parallel可以占据推理市场5-20%的份额，营收可能达到数十亿美元，并在几年内市值达到1000亿美元。 关于代理防护栏，帕拉格区分了强化学习（RL）期间的黑客行为（他认为这是一个可通过环境设置解决的问题）和恶意行为者在对齐后滥用。他认为一些被大肆宣扬的黑客行为是令人尴尬的事件，表明对策不足。他担心贫富差距和社会适应快速发展的人工智能的能力，这使得未来几年“非常非常艰难”。 帕拉格“最令人毛骨悚然的预测”是，尽管当前社会存在怀疑，但为每个人运行的常驻、持久代理的概念将变得司空见惯。他认为，虽然垂直整合有其优点，但过于僵化的公司在快速变化的科技格局中，有可能将自己排除在外。他钦佩埃隆·马斯克压缩时间并设定不合理预期的能力，推动人们超越他们自认为的能力。他提到他自己观念的转变，从仅仅专注于产品和技术，转变为认识到称职的销售和营销的重大影响。

Harry Stebbings of 20VC hosts Parag Agrawal, founder of Parallel, for a discussion on agentic search, the future of AI, and evolving business models, published on 2026-09-26. Parag introduces Parallel as "the Google for agents," providing web search capabilities specifically designed for AI agents. He emphasizes that agents will use the web 1,000 times more than humans, necessitating entirely new technology and business models. Unlike humans who type short queries and expect 10 blue links, agents require different inputs (full sentences), varying time constraints (100ms for voice agents, or no time constraint for background tasks), and different outputs (tokens or files). Parallel optimizes compute allocation in web search to save compute on the model, tailoring responses for models like cheap Luna or expensive Fable. The company's API allows agents to specify their needs (e.g., high accuracy, low latency), offering modes like "Turbo" for speed and "Advanced" for deep, expensive queries. Parallel's primary use cases span broad knowledge work, including coding, AI lawyers, insurance underwriters, and scientists. While coding uses less web search, fields like law and sales are heavily web-search oriented. Parag believes that as models become better and cheaper, enabling more agent use cases, Parallel's business thrives because they bet on agents being the web's future consumers. Discussing model trends, Parag predicts frontier models will continue to grow larger and more capable, while smaller models will achieve fixed performance levels more efficiently. He challenges the notion that 90% of token activity will go through open models and 90% of dollars through frontier models, emphasizing the value of American open models and competition. He sees current value in the model routing layer due to GPU supply/demand imbalances. Looking 3-5 years ahead, Parag identifies data and unique insights as incredibly valuable, highlighting the challenge of monetizing them for agent use (e.g., PitchBook data for venture capitalists' agents). He believes companies like Amazon, which initially resist agents, will eventually have to let them in, but the terms of engagement will be critical. A significant concern for Parag is the dissolution of the advertising industry as agents become primary customers. Since agents don't see ads, he proposes "AdSense for agents," where Parallel pays content owners a variable amount for the benefit agents derive from their content, aligning incentives and transforming web consumption from "pull to push." This, he argues, will create a larger overall market for content. Parallel operates as an infrastructure business, striving for superior quality, lower cost, and faster latency. Parag asserts that current web search pricing is misaligned with agent needs, often 50 times more expensive than necessary. He aims for Parallel to deliver web search at 1/10th the current market price, with potential for another 10x reduction, believing a "race to the bottom" on price is necessary for 1,000x scale. He explains that Parallel's technology can achieve the same quality for 1/20th to 1/50th of the compute cost of traditional human-centric web search. Parag projects Parallel could capture 5-20% of the inference market, potentially reaching billions in revenue and a $100 billion valuation within a few years. Regarding agent guardrails, Parag distinguishes between hacking during reinforcement learning (RL), which he believes is a solvable issue of environment setup, and post-alignment misuse by malicious actors. He views some celebrated hacks as embarrassments, indicating insufficient countermeasures. He worries about wealth disparity and society's ability to adapt to rapidly advancing AI, making the next few years "really, really rough." Parag's "spookiest prediction" is that the concept of always-on, persistent agents running for everyone will become commonplace, despite current societal skepticism. He believes that while vertical integration has merit, companies that are too rigid risk boxing themselves out in a fast-changing technological landscape. He admires Elon Musk's ability to compress time and set unreasonable expectations, pushing people to exceed their perceived capabilities. He cites a change in his own perspective, moving from an exclusive focus on product and technology to appreciating the significant impact of competent sales and marketing.