ExpertsFMParallel’s Parag Agrawal: Building a New Web for AI Agents
Training Data · August 25, 2026
5 Selected Lessons:
heuristic · 3 framework · 1 reasoning_pattern · 1
Agentic search architecture: layered query enrichment, multi-index retrieval, and progressive ranking
Source
Training Data — Parallel’s Parag Agrawal: Building a New Web for AI Agents (August 25, 2026) · at 10:17 ▸ · at 19:38 ▸
When this applies
when building any system that must filter an enormous candidate set down to a tiny, highest-value output under strict cost/latency budgets
The takeaway
Use a funnel of increasingly expensive, increasingly precise filtering stages so that compute is concentrated only on the shrinking set of plausible candidates, not spread evenly across the whole search space.
Why it works
By layering query understanding, multiple specialized indexes (fresh, structured/knowledge-graph, big general index), and successively larger ranking models, compute is spent efficiently — cheap filters eliminate most candidates before expensive, precise models evaluate the remainder.
One level up
Any extreme-scale narrowing problem (billions to few) should be solved via progressive, cost-aware filtering stages, not a single monolithic model.
Where it stops holding
This stops working when the cost of maintaining multiple specialized indexes and layered models exceeds the value gained from the marginal precision improvement, or when latency budgets shrink below what a multi-stage funnel can support.
Detail
procedure:
- Run initial models to interpret and enrich the incoming query
- Route the enriched query to multiple specialized indexes (big general index, fresh index, structured/knowledge-graph index)
- Apply a retrieval layer per index to narrow hundreds of billions of documents to thousands/tens of thousands
- Apply successive ranking layers with progressively bigger, more feature-rich models
- Allocate different compute/latency budgets at each stage depending on the API tier (e.g., 200ms vs 3s)
benchmarks: {"latency_improvement":"from three seconds down to 200 milliseconds","final_output_size":"about a thousand tokens from a trillion web pages"}
rationale: Every query is effectively 'give me a thousand tokens from a trillion web pages... and make sure they're the right thousand tokens.'
In their words
Like that's the query that we are getting effectively from to our search engine. And so now what you have to do, and until last week, we would give ourselves three seconds to throw a bunch of compute at read time to do it. This week, last week, we shipped a product which now does it in 200 milliseconds, right?
Draw on analogous market history (ad auctions) to predict eventual incentive-alignment adoption
Source
Training Data — Parallel’s Parag Agrawal: Building a New Web for AI Agents (August 25, 2026) · at 44:42 ▸
When this applies
when predicting whether a new, theoretically sound but initially unfamiliar market mechanism will be adopted by skeptical participants
The takeaway
Expect adoption of a complex, trust-requiring market mechanism to follow the same arc as historical analogues: initial skepticism, gradual establishment of feedback loops, then broad reliance on the mechanism once it demonstrably works, even without full participant understanding.
Why it works
A previous market (ad auctions) went through the same arc: initially unfamiliar, bid-based mechanisms were distrusted, but once feedback loops (ROI measurement) were established, the market converged on trusting automated mechanisms (auto-bid) that participants didn't fully understand mechanically but could verify delivered value.
One level up
Use historical analogues of market mechanism adoption to forecast the adoption curve of a new mechanism, rather than assuming either instant trust or permanent rejection.
Where it stops holding
This stops working when the new market's participants are far more numerous, fragmented, or lower-trust than the analogous historical market, making feedback-loop establishment much slower or impossible.
Detail
approach: Compare the current unfamiliarity of Shapley-based content pricing to the historical unfamiliarity of second-price ad auctions, and project a similar trust-building trajectory.
signals:
- second-price auctions initially required participants to trust unfamiliar bidding mechanics
- markets eventually moved to auto-bid once ROI measurement feedback loops existed
- agent traffic is growing an order of magnitude year on year
rationale: 'I think a solid foundation based on incentive alignment that Shapley Matt drives will ultimately maximize participation of content owners into this as well as content seekers via AI in an optimized system.'
In their words
I think a solid foundation based on incentive alignment that Shapley Matt drives will ultimately maximize participation of content owners into this as well as content seekers via AI in an optimized system.
Treat human feedback data as a legacy bug, not a requirement, when a better feedback source exists
Source
Training Data — Parallel’s Parag Agrawal: Building a New Web for AI Agents (August 25, 2026) · at 0:00 ▸ · at 5:31 ▸
When this applies
when an industry's standard feedback mechanism is a historical artifact of past constraints, and a cheaper, more scalable, more accurate feedback source has recently become available
The takeaway
Question whether your industry's 'default' data source is actually optimal or just a historical accident, and replace it when a new technology makes a better signal economically accessible.
Why it works
Click data was the best proxy available when only humans used the product; now models can generate expert-quality ratings cheaply and at scale, making it a strictly better signal for training ranking systems.
One level up
Technological shifts often invalidate not just products but the data collection assumptions underlying an entire industry — re-examine both.
Where it stops holding
This stops working when the new feedback source (e.g., agent/model ratings) doesn't actually correlate with real end-user value, producing a feedback loop optimized for the wrong target.
Detail
rule: Replace legacy feedback signals (like human clicks) with cheaper, more scalable signals (like model/agent-generated ratings) once the technology exists to generate them at scale.
threshold: when human ratings/click data become the binding cost/scale constraint on improving a ranking system
rationale: Models are 'really good at compressing information' and can serve as cheap expert raters, removing the historical bottleneck that only giants like Google/Bing could overcome.
In their words
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.
Sequence optimization: solve quality and cost before latency
Source
Training Data — Parallel’s Parag Agrawal: Building a New Web for AI Agents (August 25, 2026) · at 29:22 ▸
When this applies
when optimizing any system along multiple competing dimensions (e.g., quality, cost, speed) and you can't tackle all simultaneously
The takeaway
When facing multiple optimization axes, solve the dimensions with the least mature playbook first, and defer the dimensions with well-known, applicable techniques.
Why it works
Latency/systems optimization (distilling to smaller, faster models) is a 'known art,' while achieving top quality is comparatively unsolved research; tackling the harder, less-understood problem first avoids re-work and lets known techniques be applied later with confidence.
One level up
Sequence work by the maturity of available solutions, not by perceived urgency or customer complaints.
Where it stops holding
This stops working when the deferred dimension (e.g., latency) is itself the binding constraint for adoption, making customers unwilling to wait for quality-first iteration.
Detail
rule: Ignore the dimension with a known, mechanical solution (latency via distillation) until the dimension requiring genuine research (quality) is solved.
threshold: roughly 'a couple of years' focused purely on quality/cost before investing in latency
rationale: Optimizing systems and distilling to smaller models is much more known art than the unknown research of achieving top search quality.
In their words
Because optimizing systems, distilling to smaller models, is much more known art than unknown research. So once we achieved the best quality search and search agent products at every price point, we started working on latency.
Optimize agent interfaces around the actual input characteristics of the new customer type
Source
Training Data — Parallel’s Parag Agrawal: Building a New Web for AI Agents (August 25, 2026) · at 13:23 ▸
When this applies
when a new class of customer/user has fundamentally different input behavior (patience, verbosity, precision) than the one your system was originally designed for
The takeaway
Redesign the interface contract around what your new customer type is actually capable of providing, rather than inheriting the old customer type's interface conventions.
Why it works
Humans are economical/lazy typists relying on autocomplete and short, typo-ridden queries, while agents can and do provide long, precise, well-specified queries — exploiting this difference lets the search engine do less guessing and deliver better results.
One level up
Interfaces should be re-derived from the actual capabilities/constraints of the current user, not carried over by default from the legacy user.
Where it stops holding
This stops working when the new customer type (agents) is itself trained/optimized to mimic old customer type behavior (short queries), eliminating the expected advantage.
Detail
rule: Design query/input interfaces to exploit the new user type's actual capacity for detail and precision rather than defaulting to legacy brevity conventions.
threshold: n/a — qualitative design principle
rationale: 'Fewer typos, better specified queries, perhaps longer queries, less for the search engine to guess what the agent might want... you get to solve a different class of problems.'
In their words
Humans rely on keyword search. With agents, we've had to innovate quite a lot on the what does an agent tell our search engine? The more we know, the better we can do.