Agents Are Becoming the Users: The Web, the Voice Call, and the Self-Grade

Agrawal's agent web

Parag Agrawal, speaking on the Sourcery podcast, argued that AI agents will eventually use the web 5,000 to 10,000 times more than humans, and that current web business models based on human attention and random browsing are not durable in an agent-dominated web, a shift that would reorder who the web is built for. m:…3013

Agrawal's company Parallel is building a response to that shift: a system to compute and pay content owners for agent-read content in order to keep the web open, after closing a 100 million dollar Series B last month at a 2 billion dollar valuation and growing its team from 30 to 120 employees in the last year. m:…3013

The economics Agrawal sketched are concrete: for every 10 GPUs spent on model inference, the industry will spend 1 to 2 GPUs on web search, and voice agents now require web search latency of 150 to 200 milliseconds to count as acceptable products, which puts search latency, not just model quality, at the center of agent product design. m:…3013

Pipecat's voice-agent plumbing

Pipecat's version 1.14.0 release adds services that fit directly into that agent-web picture: a Perplexity agent service for Perplexity's Agent API, an NVIDIA Omni service for Nemotron Omni models, and text transforms for rewriting output aggregations, alongside a default model update for its NVIDIA service to a Nemotron 3.5 super vision-language model. m:…8225

The release also shows the maintenance side of voice infrastructure: the older Perplexity service is deprecated in favor of the new agent service, a websocket retry method is deprecated ahead of removal in version 2.0, the speech service now supports 8 kHz audio, and the Twilio serializer now supports authentication with an API key. m:…8225

Security maintenance appears as well, with the minimum aiohttp version raised to 3.14.4 to address vulnerabilities, and a retirement note that Vertex AI will retire a Gemini flash model on November 19, 2026, a reminder that voice agents depend on upstream model availability as much as on their own code. m:…8225

Agents that grade themselves

A preprint benchmark called EvalResearchBench asks whether agents can automate benchmark design to reduce evaluation costs, testing researcher agents that select or synthesize tasks, implement graders, and revise them in pilot tests before freezing an executable evaluator, across 9 researchers and 13 candidate models compared against 14 target benchmarks. m:…4198

The preprint's results are mixed in a useful way: generated evaluators currently fall short of target agreement, the best evaluators order about 75 percent of candidate pairs as the targets do against a ceiling of 91 percent agreement among target benchmarks themselves, and the best evaluator on development targets is not the best on sealed targets hidden from the researcher. m:…4198

The failure modes are specific rather than vague: generated evaluators suffer from budget exhaustion and score dominance, evaluators can truncate answers, exhaust the evaluation budget, or let a few questions dominate a domain score, yet agents can also repair tasks and graders through pilot feedback, and the evaluator with the lowest recorded execution cost attained the highest pairwise agreement. m:…4198

One user class, three layers

The agent story has three layers that fit together: Agrawal's web that agents will use thousands of times more than humans, Pipecat's voice plumbing that lets agents talk and search within tight latency budgets, and EvalResearchBench's finding that agents can build evaluators but still fall short of target agreement, which means the agent web needs both infrastructure and better self-measurement. m:…3013 m:…8225 m:…4198

Payment and latency are the practical constraints: Parallel's plan to compute and pay content owners for agent-read content addresses who gets paid when agents read, while the 150 to 200 millisecond search latency bar for voice agents and the budget-exhaustion failures of generated evaluators show where agent products still break. m:…3013 m:…4198