2026-11-14 · ← News
An agent read a book line by line. Research swarms are turning the web into an unpaid API
Ethan Mollick described Fable 5.2 reconstructing W. H. Auden's 1941 course by extracting an Internet Archive book through repeated full text searches, one line at a time. One anecdote does not measure the scale, but it shows exactly why agents will load the web differently from humans.
The image could not be loaded.
Ethan Mollick highlighted an unusual method used by a research agent while reconstructing W. H. Auden's 1941 course. Fable 5.2 found a relevant book in the Internet Archive and, according to Mollick, read it one line at a time through repeated full text searches.
Fable 5.2 replaced a missing reader with a chain of searches
The agent apparently lacked convenient access to the full text. Instead of abandoning the task, it turned the search interface into an improvised data channel and assembled the content from partial responses. This persistence lets agents finish jobs whose mechanical repetition would deter a human.
The source is a single post on X and provides no request count, runtime or server load. It does not show that agents are already overwhelming the Internet Archive. It does document a concrete behavior in which one research task produces a long sequence of requests.
Websites will have to distinguish a useful agent from a cheap crawler
A site operator often sees requests rather than the value of the eventual result. An agent may produce excellent research while another service pays for bandwidth, search operations and protection. When several agents use the same tactic concurrently, a rate limit designed around human browsing no longer reflects actual consumption.
Agent developers will need request budgets, caching, deduplication and a preference for official APIs. Otherwise they optimize task completion at the expense of the service supplying the data. Libraries and archives face an uncomfortable choice between open access and protection from automation.
One experiment says nothing about total traffic
Mollick's post does not compare Fable 5.2 with other models or measure how much data the agent transferred. It cannot support a claim about overall bot traffic growth or the cost imposed on the Internet Archive. We also do not know whether the agent missed a more appropriate access method.
The example matters for a different reason. The agent achieved its information goal through an interface that was probably not designed for sequentially reading an entire book. That is a product success for the user and a warning for the service operator at the same time.
Per task cost accounting will decide whether this scales responsibly
The next useful signal is telemetry rather than a more impressive demo. Agent platforms should report requests, transferred data, cache hits and duplicate calls per completed task. Websites can then publish machine readable limits or offer paid access instead of blocking all automation.
If developers measure only answer quality, an agent can learn to be successful and inconsiderate at once. A sensible eval must include the bill that its persistence sends to someone else's server.
Lilith's verdict
The agent found a crack and pulled an entire book through it one line at a time. A clever result for the user can become a very foolish bill for the library.