A personal assistant repurposed for data collection
A developer's account of using Meta's Muse assistant suggests that browser-based agents can perform large web-research jobs at low apparent marginal cost, while creating a new challenge for sites trying to distinguish useful automation from abusive scraping. The report is an individual experience rather than an independent benchmark, but it documents a concrete workflow and the concerns it raised for its author.
The developer used Muse to support fullsets.fm, a project that gathers professionally recorded concert videos, groups them by artist and extracts track lists. The main discovery problem was finding suitable performances scattered across small YouTube channels. Muse was instructed to search Reddit and YouTube for large sets of candidate videos and to assemble artist lists before checking performers one by one.
Muse also operated later in the pipeline, reviewing outputs from other AI models and helping decide whether a video met the project's criteria. The user said an $80 monthly subscription supported a few billion tokens of weekly work, sometimes without token use appearing to be counted, along with access to a managed cloud machine running Chrome. Those figures are self-reported and should not be read as a guaranteed service allowance or price-performance result.
Scale creates a defensive problem
The account's larger significance lies in how easily a consumer agent can turn instructions into sustained browser activity. A cloud-hosted process does not need breaks, and multiple jobs could potentially run in parallel. From the user's perspective, that makes tedious discovery work practical. From a website operator's perspective, the same capability can impose server load, copy large quantities of information or mimic ordinary browsing closely enough to complicate enforcement.
The author expressed concern about the effect if similar agents targeted his own websites and predicted more services would block Muse unless Meta constrained misuse. He also said Amazon had already blocked the service, though the supplied evidence contains no statement from Amazon explaining its action. The experiment does not establish whether the described collection complied with every site's terms or robots policies.
Muse was not presented as universally capable. The user described its Spark 1.3 model as unsuitable for his writing or programming needs and reported that jobs sometimes froze. Communication with the agent and managing several browser tasks through the interface could also be frustrating. The value came from high-volume, comparatively simple classification and browsing work rather than from polished generation.
This mix of utility and roughness illustrates a likely pressure point for agent platforms. Subscription products designed to help a person complete ordinary online tasks can also become inexpensive scraping infrastructure when they include a browser, ample compute and extensive model usage. Providers may need rate limits, clearer acceptable-use enforcement and mechanisms for websites to identify automated traffic. Publishers and platforms, meanwhile, will have to decide which agents to admit and under what conditions. One developer's project cannot measure the network-wide impact, but it shows that this debate is already moving from hypothetical autonomous browsing to persistent, user-directed collection work.
