45% of AI Budget Goes to Always-On Agents, With Anthropic and OpenAI Products Called Out
Dylan Patel revealed on X that 45% of his team's AI spending already goes to persistent agents. He called out Claude Tags and OpenAI's Slack agent as hard to use, saying they mainly use Perplexity Computer for a practical reason: feature requests can land within 24 hours.
Dylan Patel posted on X that 45% of his team's AI spending has already gone to persistent agents. He calls this kind of thing the next unlock for AI use. For non-coding readers, think of it as an AI colleague that stays running in the background: unlike one-off Q&A, it runs continuously, takes on tasks, checks status, and handles workflows.
He then called out the product experience. Claude Tags is hard to use, and OpenAI's Slack agent is even worse. The team mainly uses Perplexity Computer, which calls multiple models; Perplexity's Harness is very strong. The key takeaway here is that model capability and product capability are being evaluated separately. A product can mainly run Anthropic's models, but users may not be willing to use the shell Anthropic built itself.
In replies, Dylan put it more bluntly: Perplexity mainly uses Anthropic models, but Anthropic's product is not good and does not take product feedback well; the same goes for OpenAI. Conversely, Perplexity is fast on feature requests, and needs can be implemented within 24 hours. He even mentioned that Perplexity took the name from them in February.
In the comments, you can see real usage of always-on agents. One user said Grok Bot alone accounted for 14% of the Grok quota last week, while chat accounted for only 1%; the always-on one will eat up the quota. Others thought 45% was exaggerated—half of their own agents were just checking other agents. Others asked what the remaining 55% is if 45% goes to persistent agents, and what exactly is wrong with Claude Tags.
Taken together, these reactions make the competitive battleground for always-on agents very concrete: orchestration, quotas, feedback speed, and the product shell. Models are the underlying capability; what users touch every day is another layer. Dylan's tweet stands out because it pierces a thin veil: model providers may not be able to win the product experience just by relying on their own models, while fast-following product teams may instead be the ones to capture the demand.
Original post: https://x.com/dylan522p/status/2120617704101155270
发布时间: 2026-11-12 18:26