How Web Data Infrastructure Powers the Next Generation of AI — Patricija Žemaitytė, Oxylabs
Fresh web retrieval becomes product infrastructure once agents need sub-second responses and sustained scale. The talk’s failed client demo shows why dev latency alone says little about production resilience.
Oxylabs rebuilt a roughly four-second search path for an **under-800 ms** requirement and reached about **650 ms P90** in under two weeks, then got blocked during the client demo. Another system was pushed from 10,000 to **60,000 requests per second** in under two months.
Treat live web access as a production dependency: minimize the response surface, test realistic traffic, instrument end-to-end jobs and design around browser, proxy, parsing and retry costs. For interactive agents, latency determines whether retrieval can remain inside the action loop.
Oxylabs rebuilt a roughly four-second search path for an **under-800 ms** requirement and reached about **650 ms P90** in under two weeks, then got blocked during the client demo. Another system was pushed from 10,000 to **60,000 requests per second** in under two months. Treat live web access as a production dependency: minimize the response surface, test realistic traffic, instrument end-to-end jobs and design around browser, proxy, parsing and retry costs. For interactive agents, latency determines whether retrieval can remain inside the action loop. These are vendor case studies, not controlled benchmarks. The talk also says synthetic load was easier than representative organic traffic, and its current target of about 100,000 requests per second remains ongoing rather than a completed result.
This shifts web retrieval from a supporting capability to a latency- and reliability-critical part of the agent action loop. Against the prior observability and serving work, it broadens performance accounting beyond model and gateway metrics to browser, proxy, parsing, retry, and representative-traffic costs. The vendor cases suggest substantial optimization headroom, but their workload-specific and partly unfinished evidence prevents treating the reported latency or throughput as general targets.