Sequoia Capital · Published 2026-08-25

Parallel’s Parag Agrawal: Building a New Web for AI Agents

privateReview: unreviewedConfidence: high
agentic searchIncremental indexingShapley valuesincentive alignmentweb infrastructureagentic-searchincremental-indexingretrieval-architectureagent-feedbackshapley-valuesincentive-alignmentcontext-windowdifferential-pricing
Open on YouTube ↗

Summary

Overview

  • Speaker: Parag Agrawal
  • Channel: Sequoia Capital
  • Main topic: Building agentic search and web infrastructure for AI agents
  • Purpose: To explain the vision behind Parallel Web Systems, the economics of agentic web search, and the fundamental differences between human-driven and agent-driven information retrieval. Parag Agrawal, former CEO of Twitter and founder of Parallel Web Systems, discusses the evolution of web search from human-centric browsers to agent-centric search engines. He explains why human click data is a bug for AI agents, how incremental indexing and search agent architecture enable cost-effective scale, and why incentive alignment between content owners, seekers, and AI models is crucial for the future of the internet.

Topic Map

Human Click Data vs. Agent Feedback

  • Explanation: Human feedback loops are based on latency-optimized, click-driven heuristics designed for lazy humans, whereas AI agents require objective agent feedback and compressed information.
  • Key claims:
    • Human click data is a bug for agentic search.
    • Agents need agent feedback, not human feedback.
    • Indexes built for human latency are suboptimal for agents.
  • Examples:
    • Skimming Google search results pages 1 to 7 out of laziness.
  • Terminology:
    • agent feedback
    • human click data
    • latency optimization
  • Why it matters: Redefines how search engines must be engineered when the primary consumers are autonomous software agents rather than human users.

The Problem of Web Search at Scale

  • Explanation: Traditional web search requires crawling trillions of pages, building massive indexes, and running expensive matching algorithms for every query.
  • Key claims:
    • Web search is a billion-to-billion matching problem.
    • Crawling the entire web upfront is extremely expensive.
    • Indexes act as latency optimizations for humans.
  • Examples:
    • Matching 1,000 tokens of query against a trillion web pages.
  • Terminology:
    • crawling
    • indexing
    • retrieval layer
    • ranking layer
  • Why it matters: Illustrates why traditional search infrastructure does not scale economically for high-frequency agentic workloads.

Parallel's Incremental Indexing Approach

  • Explanation: Instead of crawling everything upfront, Parallel builds search agents that retrieve relevant excerpts and bring them into the agent's context window incrementally.
  • Key claims:
    • Parallel started by building search agents before search products.
    • Incremental index building reduces upfront capital expenditure.
    • Agent-driven search relies on precise context window feeding.
  • Examples:
    • Insurance underwriting workflows and financial data enrichment.
  • Terminology:
    • search agent
    • incremental indexing
    • context window
  • Why it matters: Offers a viable economic path for startups to compete with tech giants in web indexing.

Incentive Alignment and Shapley Values

  • Explanation: Ensuring content owners get paid fairly when their data is used by AI agents, utilizing game-theoretic concepts like Shapley values.
  • Key claims:
    • Fixed-price contracts in AI inference do not drive long-term sustainability.
    • Incentive alignment maximizes participation of content owners and seekers.
    • Shapley values help distribute collaborative value mathematically.
  • Examples:
    • Comparing a banker reading a report vs. a retired dad reading financial data.
  • Terminology:
    • Shapley values
    • incentive alignment
    • differential pricing
    • game theory
  • Why it matters: Addresses the economic sustainability of the web in an era where AI models ingest content without driving traditional ad clicks.

Key Points

Agents require agentic search

  • Explanation: AI agents operating 1,000x more frequently than humans need search infrastructure optimized for token efficiency and programmatic precision rather than human browsing habits.
  • Evidence: Web traffic from AI is scaling rapidly, matching or exceeding human page reads in specific domains.
  • Practical implication: Developers building multi-agent systems must integrate search APIs designed specifically for programmatic agent consumption.

Incremental building over upfront capital expenditure

  • Explanation: By deploying search agents that fetch data on demand and incrementally build indexes, companies can avoid the prohibitive costs of crawling the entire web upfront.
  • Evidence: Parallel's journey from building search agents for specific workflows to scaling a web-wide index.
  • Practical implication: Startups can build competitive search infrastructure without multi-billion-dollar data center investments.

The collapse of traditional advertising business models

  • Explanation: Ad-supported web business models rely on human attention and click-throughs, which break down when AI agents consume content directly.
  • Evidence: Publishers signing licensing deals with AI labs because traditional ad revenue streams are threatened.
  • Practical implication: Content creators and platform architects must adopt differential pricing and incentive alignment frameworks.

Frameworks, Models & Processes

Shapley Values for Content Valuation

  • How it works: A game-theoretic approach to determine how much individual contributors (content owners) add to a collaborative outcome (AI inference result) when the whole is greater than the sum of parts.
  • Components:
    • Collaborative coalitions
    • Marginal contribution calculation
    • Differential pricing based on data uniqueness
  • When to use: When distributing revenue or compensation among multiple content sources that feed an LLM inference pipeline.

Examples & Case Studies

Building early search agents for insurance underwriting and financial workflows.

  • Illustrates: How Parallel tested search agents on specialized internal data before scaling to the web.
  • Lesson: Start with high-value vertical use cases where expert feedback is cheap and domain-specific.

Transitioning from Chaple Inc to Parallel Web Systems.

  • Illustrates: The evolution of naming and rebranding a B2B AI infrastructure company.
  • Lesson: Terrible initial names can be pivoted into strong brands once the product matures.

Actionable Takeaways

  • Immediate:
    • Recognize that human-centric web search is insufficient for autonomous AI agents.
    • Adopt incremental indexing strategies to reduce infrastructure costs.
  • Strategic:
    • Align economic incentives between content creators, AI platforms, and end users.
    • Prepare for the shift from pull-based web browsing to push-based agentic workflows.
  • Questions to investigate:
    • How will copyright and content licensing evolve as agentic search scales?
    • What is the exact tipping point where incremental indexing matches full-web crawl quality?

Claims Worth Verifying

  • AI web traffic is approaching parity with human page reads on certain platforms. (empirical)
  • Shapley values can effectively solve attribution and monetization for LLM training data. (theoretical)

Notable Quotes

"our view at parallel is that human click data is a bug and agent doing work with search should rely on agent feedback not human feedback." (at 0:00) "Shapley values is this very theoretical mathematical concept." (at 70:32)

Compressed Summary

  • Human click data is suboptimal for AI agents which require precise agent feedback.
  • Parallel builds incremental search infrastructure to avoid massive upfront web crawling costs.
  • Incentive alignment and Shapley values provide a mathematical foundation for content monetization in the AI era.
  • Agentic search shifts the web from pull-based browsing to push-based execution.
  • Keywords: agentic search, incremental indexing, shapley values, incentive alignment, web infrastructure
  • Core insight: As AI agents replace humans as the primary consumers of web data, search engines and business models must transition from human-centric latency optimization to programmatic agent-centric efficiency and fair incentive alignment.

Core insights

6
Mental Modelhigh noveltymoderate evidence

Signals derived from human behavior (clicks, latency-driven skimming) are actively harmful as training/ranking signal for agent-facing search: human click data is treated as a bug, and the correct signal is 'agent feedback' — outcome/objective signals from the agent that consumes the results.

Why it matters

If you copy a human-facing retrieval stack into an agentic system, you inherit a reward signal optimized for the wrong consumer. The feedback loop, not the index, becomes the thing that must be redesigned.

Generalization

Any system where the proxy signal (user clicks, dwell time, thumbs) was optimized for a different consumer class than the one now driving traffic will silently misalign; instrument the loop for the actual consumer's objective.

our view at parallel is that human click data is a bug and agent doing work with search
Open source video
Agents need agent feedback, not human feedback.
Open source video
Indexes built for human latency are suboptimal for agents.
Open source video
Architecturehigh noveltystrong evidence

Invert the classic search architecture: instead of crawling and indexing the whole web up front, deploy search agents that fetch relevant excerpts on demand and incrementally materialize them into the consuming agent's context window, building the index as a side effect of usage.

Why it matters

Turns a fixed multi-billion-dollar capex crawl into a usage-proportional opex, and makes the context window the effective boundary of the retrieval index — a materially different cost and quality curve for anyone building retrieval for agents.

Generalization

Lazy/on-demand materialization of a derived index (build it when queried, not before) is a viable strategy for any derived-artifact system where full precomputation is economically prohibitive.

Instead of crawling everything upfront, Parallel builds search agents that retrieve relevant excerpts and bring them into the agent's context window incrementally.
Open source video
Incremental index building reduces upfront capital expenditure.
Open source video
Parallel started by building search agents before search products.
Open source video
Predictionmedium noveltyweak evidence

Agent traffic is claimed to run roughly 1,000x more frequently than human traffic, which means the optimization target for retrieval shifts from human-perceived latency to token efficiency and programmatic precision.

Why it matters

Sizing, caching, batching, and cost models built around human query volumes (with human latency budgets) will be off by orders of magnitude; the design constraint becomes cost-per-token-of-useful-context rather than cost-per-pageview.

Generalization

When the request rate of a new consumer class exceeds the incumbent's by orders of magnitude, the incumbent's optimization target (latency, UX) must be replaced by the new class's target (throughput, token/cost efficiency, machine precision).

AI agents operating 1,000x more frequently than humans need search infrastructure optimized for token efficiency and programmatic precision rather than human browsing habits.
Open source video
Web traffic from AI is scaling rapidly, matching or exceeding human page reads in specific domains.
Open source video
Mechanismhigh noveltyweak evidence

Shapley values are proposed as the mechanism for valuing each content source's marginal contribution to an AI inference result, with differential pricing tied to data uniqueness rather than flat fixed-price contracts.

Why it matters

Gives a concrete (if unproven) computational answer to the attribution problem that otherwise blocks content owners from participating at all — attribution is the precondition for the data supply chain existing.

Generalization

Credit assignment among many contributing sources to a single composite output is a general problem (RAG over many documents, multi-source training data, multi-tool agent trajectories) where coalition-based marginal-contribution schemes are one candidate formalization.

Shapley values help distribute collaborative value mathematically.
Open source video
Fixed-price contracts in AI inference do not drive long-term sustainability.
Open source video
Differential pricing based on data uniqueness
Open source video
Failure Modemedium noveltymoderate evidence

Ad-supported web economics are a load-bearing dependency that fails when agents consume content without generating clicks; the failure is mechanical (attention → revenue chain broken), not a matter of publisher preference.

Why it matters

An agent architecture that reads the web at scale silently destroys the economic substrate of the content it depends on — a systemic reliability risk for any long-lived agentic retrieval product, not just a business-model concern.

Generalization

Any agent pipeline that extracts value from a third party without generating the third party's monetization signal creates a slow-burning supply-side failure; design the incentive channel alongside the technical one.

Ad-supported web business models rely on human attention and click-throughs, which break down when AI agents consume content directly.
Open source video
Publishers signing licensing deals with AI labs because traditional ad revenue streams are threatened.
Open source video
Practicemedium noveltymoderate evidence

Sequencing matters: build search agents for narrow, high-value vertical workflows (insurance underwriting, financial data enrichment) where expert feedback is cheap and domain-specific, then generalize to a web-wide index.

Why it matters

Gives a concrete bootstrapping path for teams that cannot afford a web-scale crawl, and explains why domain feedback loops precede general index quality.

Generalization

When the general-purpose artifact is economically unreachable at the start, bootstrap the retrieval/feedback machinery on a domain where ground truth is cheap, then widen scope.

Start with high-value vertical use cases where expert feedback is cheap and domain-specific.
Open source video
How Parallel tested search agents on specialized internal data before scaling to the web.
Open source video

Deep dives

5

Agent feedback as a replacement for click signals in retrieval ranking

Research question

If the consuming agent's task outcome (success/failure, downstream tool-call correctness, answer acceptance) is used as the ranking reward instead of human clicks, does retrieval quality on agent tasks measurably diverge from click-trained rankers — and in which direction?

Why

The claim that 'human click data is a bug' is an assertion without an operational definition of the replacement signal. Until 'agent feedback' has a collection mechanism, a noise model, and a measurable effect on agent task success, every downstream architectural decision (evaluation harnesses, ranker training, context assembly) rests on an unvalidated premise.

our view at parallel is that human click data is a bug and agent doing work with search
Open source video
Agents need agent feedback, not human feedback.
Open source video
Indexes built for human latency are suboptimal for agents.
Open source video
Source video

The crossover point where incrementally built indexes match full-web crawl quality

Research question

As a function of query volume, query skew, and corpus size, at what point does on-demand incremental index materialization reach parity (recall/precision) with a prebuilt full-crawl index, and at what cost per query?

Why

Incremental indexing is the load-bearing economic claim ('reduces upfront capital expenditure') that lets a startup compete with web-scale incumbents. If parity only arrives at volumes nobody reaches, the strategy is a cost deferral, not a substitute architecture. This is explicitly flagged as the source's own open question and has no published benchmark.

Instead of crawling everything upfront, Parallel builds search agents that retrieve relevant excerpts and bring them into the agent's context window incrementally.
Open source video
Incremental index building reduces upfront capital expenditure.
Open source video
How Parallel tested search agents on specialized internal data before scaling to the web.
Open source video
Source video

Tractability and gameability of Shapley-based content attribution at corpus scale

Research question

Can sampled or truncated Shapley approximations produce stable, non-gameable per-source payments across realistic content-coalition sizes, or does the approximation error swamp the payment signal that content owners would need to trust?

Why

Attribution is the precondition for the content supply chain existing at all under agentic consumption. But the summary presents Shapley values as mathematical theory with no demonstrated tractability, and a payment scheme that can be gamed by strategically-duplicated or subsidized content is worse than a flat contract because it invites adversarial participation.

Shapley values help distribute collaborative value mathematically.
Open source video
Fixed-price contracts in AI inference do not drive long-term sustainability.
Open source video
Differential pricing based on data uniqueness
Open source video
Source video

Measuring agent-vs-human traffic crossover with server-side classification

Research question

Across a broad domain panel, what fraction of page reads originates from AI agents, how is that traffic identifiable server-side, and how much does the measured number move depending on the classification method (user-agent, protocol signature, request shape)?

Why

The 1,000x frequency multiplier and the parity claim are the empirical foundation for the entire argument that infrastructure must be redesigned. Both are qualified to 'specific domains' with no data shown. Without a reproducible measurement method, the premise cannot be confirmed or falsified, and any capacity/cost model built on it is unfounded.

AI agents operating 1,000x more frequently than humans need search infrastructure optimized for token efficiency and programmatic precision rather than human browsing habits.
Open source video
Web traffic from AI is scaling rapidly, matching or exceeding human page reads in specific domains.
Open source video
Source video

The mechanical breakdown of attention-mediated revenue under agentic consumption

Research question

For a given publisher, how much does agent-originated consumption of the same content reduce click revenue per unit of content served, and does a per-contribution licensing channel recover it at a rate that keeps the supply side producing?

Why

This is a systemic reliability risk, not a business-model curiosity: an agentic retrieval product that reads the web at scale degrades the substrate it depends on. Treating it as an incentive-channel design problem alongside the technical one is the only way to avoid building on a supply side that quietly exits.

Ad-supported web business models rely on human attention and click-throughs, which break down when AI agents consume content directly.
Open source video
Publishers signing licensing deals with AI labs because traditional ad revenue streams are threatened.
Open source video
Source video

Article ideas

4

Your Retrieval Stack Is Optimized for a Consumer That No Longer Exists

The most consequential bug in agentic search is not the index but the feedback loop: a ranker trained on human clicks is a ranker optimized for skimming behavior, and porting it into an agent-facing system import a reward signal that actively rewards the wrong thing. The fix is not a better crawler but an evaluation harness built around agent task outcomes.

Angle

Treat the feedback loop as the migration artifact, not the index. Show what concretely breaks when CTR-shaped metrics are the only labels you have, and what must be instrumented instead.

Source video

Build the Index You Actually Query: Lazy Materialization as an Economic Strategy

Precomputing a derived artifact for the whole universe of possible inputs is a capex decision that only incumbents can afford; materializing it on demand and caching the byproduct of use is a legitimate substitute, not a degraded fallback — but only up to a coverage threshold nobody has measured. The honest version of this argument states the threshold as an open question, not a solved one.

Angle

Cost-curve framing: compare usage-proportional retrieval cost against amortized full-crawl cost as query volume and skew vary, and make the crossover explicit rather than assumed.

Source video

Agents Read the Web for Free; Someone Still Has to Pay for It

Ad-supported content is a load-bearing dependency of every agentic retrieval product, and its failure under agent consumption is mechanical rather than a matter of publisher preference. Attribution and provenance are therefore runtime requirements of the retrieval pipeline, not finance-side concerns — a system that cannot account for per-source contribution cannot participate in the market that keeps its inputs alive.

Angle

Trace the broken chain (attention to clicks to revenue) and derive what the pipeline must carry: per-excerpt provenance, per-source contribution accounting, and a pricing channel designed alongside the technical one.

Source video

Bootstrap the Loop Before You Build the Index

When the general-purpose artifact is economically unreachable at the start, the right first move is not a smaller crawl but a higher-value domain: insurance underwriting and financial data enrichment supply cheap expert ground truth, which is what makes an agent-feedback loop trainable before web-scale index coverage exists. Sequencing, not architecture, is the binding constraint for teams without incumbent capital.

Angle

Contrast the vertical-bootstrap path against the general-crawl path on feedback cost per labeled outcome, and show why ground-truth availability — not corpus size — gates early progress.

Source video

Project ideas

4

Agent-Outcome Ranker Harness

beyond-evals

A ranker trained and selected on agent task-outcome labels will beat a click-trained ranker on agent task success rate, while scoring worse on human-behavior proxy metrics — meaning the two objectives are measurably divergent rather than correlated.

Proof of concept

Stand up a dual-objective retrieval loop over a fixed document set: two rankers (one trained on click-style proxy labels, one on downstream task-outcome labels), a scripted agent consumer that executes a task using retrieved excerpts, and automatic success grading. Run both rankers against an identical held-out agent task suite.

Measurement

Delta in agent task success rate between rankers, plus the correlation coefficient between agent-outcome scores and click-proxy scores across the ranked lists. A near-zero or negative correlation confirms divergence; a high positive correlation falsifies the premise that human click data is a bug.

Source video

Provenance Gate and Contribution Ledger

gatehouse

Attaching per-excerpt provenance metadata at retrieval time and computing sampled per-source marginal contribution is tractable at production query rates, and the resulting payments remain stable when an adversarial source injects near-duplicate content.

Proof of concept

Instrument a retrieval pipeline so every excerpt carries a signed source identifier that survives into the model context and output. Implement sampled/truncated Shapley approximation over the retrieved coalition per query, write contributions to an append-only ledger, then run a gaming experiment where a synthetic source floods near-duplicates and observe whether its payout inflates beyond its true marginal value.

Measurement

Approximation error versus exact Shapley on small coalitions, query-time overhead percentage attributable to provenance and attribution, and payout inflation factor for the adversarial source. Failure mode is a payout inflation factor materially above 1.0.

Source video

Incremental-vs-Full Index Crossover Benchmark

new

Incremental, query-driven index materialization reaches quality parity with a prebuilt full index on a realistic query distribution before the cumulative retrieval cost of the incremental approach exceeds the amortized cost of the prebuilt index — i.e., there exists a usable volume window where incremental wins on both axes.

Proof of concept

Build a paired benchmark over the same corpus and query set: (a) a fully prebuilt index, (b) a search agent that fetches and materializes excerpts on demand and caches them. Replay a synthetic query stream with controlled skew, sweeping query volume upward, and record recall and precision against a gold set at each volume step, alongside cumulative cost per query.

Measurement

Recall/precision curves as a function of cumulative query volume for both arms, plus cumulative cost per query. The hypothesis is falsified if the incremental arm's quality curve stays below the prebuilt arm's across the entire range where its cost is lower — i.e., no parity window exists.

Source video

Traffic Classification Panel

movement-lab

Agent traffic share is measurable server-side with low classification ambiguity, and the agent-to-human read ratio varies by more than an order of magnitude across domain categories — meaning a single global multiplier is a misleading basis for capacity planning.

Proof of concept

Deploy request-level classification (user-agent, protocol signature, request shape, session behavior) across a panel of instrumented domains spanning content categories, run the classifier alongside manual ground-truth labeling on a sample, and compute per-category agent-to-human read ratios.

Measurement

Classifier precision/recall against manually labeled ground truth, and the spread of agent-to-human read ratios across categories. A tight spread would falsify the cross-domain-variance hypothesis; a broad spread would falsify any single-multiplier planning model.

Source video

Architectural implications

3

Where the index lives and when it is built changes from a global precomputed artifact to a per-query, agent-driven materialization.

Before

Crawl the web up front, build one large index, run a retrieval + ranking layer per query; the index is a latency optimization for human users.

After

Search agents retrieve relevant excerpts on demand and incrementally feed the consuming agent's context window, building index coverage as a byproduct of use.

Consequence

Retrieval cost becomes proportional to usage rather than to corpus size, and context-window assembly becomes a first-class component that must be budgeted and evaluated.

Source video

The ranking/feedback loop changes owner and signal type.

Before

Ranking models trained and evaluated on human click behavior.

After

Ranking driven by agent feedback — objective outcome signals from the programmatic consumer.

Consequence

New evaluation harnesses are required that can capture and score agent outcomes rather than clicks; offline metrics like CTR lose meaning.

Source video

Content access becomes an economic gate rather than an open crawl.

Before

Open crawl + ad monetization; the publisher is paid indirectly by human attention.

After

Licensing deals and differential pricing keyed to the marginal value of each content source.

Consequence

Retrieval systems must carry provenance/attribution metadata per retrieved excerpt, since payment and permission depend on per-source contribution accounting.

Source video

Tradeoffs and failure modes

4

Incremental indexing vs. full-web crawl quality

Benefit

Avoids prohibitive upfront capital expenditure on crawling the entire web.

Cost or risk

Unknown quality ceiling: there is no established point at which on-demand incremental coverage matches a full crawl for a given query.

What is the exact tipping point where incremental indexing matches full-web crawl quality?
Open source video
Source video

Human-latency-optimized indexes reused for agents

Benefit

Mature infrastructure and known latency characteristics for human browsing.

Cost or risk

Structurally suboptimal for agents, which need compressed information and token efficiency rather than low-latency page presentation.

Indexes built for human latency are suboptimal for agents.
Open source video
Source video

Shapley-value attribution for content valuation

Benefit

Mathematically principled distribution of collaborative value among content owners, replacing fixed-price contracts.

Cost or risk

Attribution/monetization claims for LLM training data remain theoretical; no demonstrated tractability or adoption is provided.

Shapley values can effectively solve attribution and monetization for LLM training data. (theoretical)
Open source video
Source video

Ad-supported content supply under agentic consumption

Benefit

Free, openly crawlable content at web scale.

Cost or risk

Agents consume content without producing the clicks the model monetizes, threatening the supply side the agents depend on.

Ad-supported web business models rely on human attention and click-throughs, which break down when AI agents consume content directly.
Open source video
Source video

Open questions

5

What exactly constitutes 'agent feedback' as a training and ranking signal, and how is it collected at scale?

Why unresolved

The summary asserts agents need agent feedback rather than human feedback but does not describe the signal, its collection mechanism, or its noise characteristics.

Research direction

Prototype an outcome-labeled retrieval loop where the consuming agent's task success/failure is the reward signal, and measure how ranker quality diverges from click-trained baselines.

Source video

At what coverage or query-distribution threshold does incremental, on-demand indexing match full-web crawl quality?

Why unresolved

Raised explicitly as an open question in the source; no benchmark or measurement is given.

Research direction

Build a paired benchmark: same query set served by a full prebuilt index versus an incrementally built index, measuring recall/precision and cost per query as a function of usage volume and query skew.

Source video

Are Shapley values computationally and practically tractable for attributing value across web-scale content coalitions?

Why unresolved

Presented as a game-theoretic mechanism and flagged in the summary as a theoretical claim rather than a demonstrated result.

Research direction

Evaluate approximation schemes (sampled/truncated Shapley) on realistic content-coalition sizes and test whether resulting payments are stable and non-gameable.

Source video

How will copyright and content licensing evolve as agentic search scales?

Why unresolved

Listed as an open question in the source; the current licensing-deal pattern is described as reactive to ad revenue loss.

Research direction

Track licensing terms and per-source pricing structures as agent traffic share grows, and model the equilibrium between crawl openness and gated access.

Source video

What is the measured gap between agent web traffic and human page reads, and in which domains?

Why unresolved

The parity claim is stated as empirical but is qualified to 'certain platforms'/'specific domains' with no data shown.

Research direction

Instrument server-side traffic classification (agent vs. human user-agent/protocol signatures) across many domains to quantify the crossover point.

Source video

Key claims

7
opinionVerification needed

Human click data is a bug for agentic search; agents should be driven by agent feedback instead.

Evidence

our view at parallel is that human click data is a bug and agent doing work with search

Question

Does ranker/retrieval quality actually degrade when trained on human click signals versus agent outcome signals, measured on agent task success?

Source video
factualVerification needed

Web search is a billion-to-billion matching problem requiring crawling trillions of pages and running expensive matching per query.

Evidence

Web search is a billion-to-billion matching problem.

Question

What are the actual cost and latency profiles of full-crawl-and-index versus on-demand agent retrieval per unit of query volume?

Source video
causalVerification needed

Incremental, on-demand index building reduces upfront capital expenditure relative to crawling the entire web upfront.

Evidence

Incremental index building reduces upfront capital expenditure.

Question

At what request volume does usage-proportional retrieval cost exceed the amortized cost of a prebuilt index?

Source video
comparativeVerification needed

AI agents operate roughly 1,000x more frequently than humans and therefore need infrastructure optimized for token efficiency rather than human browsing patterns.

Evidence

AI agents operating 1,000x more frequently than humans need search infrastructure optimized for token efficiency and programmatic precision rather than human browsing habits.

Question

Is the ~1000x frequency multiplier empirically observed, and over what base of human query volume?

Source video
factualVerification needed

AI web traffic is approaching parity with human page reads on certain platforms.

Evidence

Web traffic from AI is scaling rapidly, matching or exceeding human page reads in specific domains.

Question

Which domains show agent traffic at or above human page reads, and how is agent traffic identified in the measurement?

Source video
causalVerification needed

Fixed-price contracts in AI inference do not drive long-term sustainability, whereas incentive alignment does.

Evidence

Fixed-price contracts in AI inference do not drive long-term sustainability.

Question

Do differential/contribution-based pricing schemes measurably increase content-owner participation versus flat licensing deals?

Source video
predictionVerification needed

Shapley values can effectively solve attribution and monetization for LLM training data.

Evidence

Shapley values can effectively solve attribution and monetization for LLM training data. (theoretical)

Question

Can Shapley-based attribution be approximated tractably and non-gameably at web-corpus scale?

Source video

Connections

5