Not sure anything can be done about this because it seems to be part of the training/guardrails on Anthropic's end? Every so often Claude Opus and 3.5 Sonnet will refuse to go to a provided url or website, saying that they do not have access to the internet. Sometimes it is an outright refusal to follow the prompt, sometimes, it uses the the ResearchAgent anyway while saying it cannot. So far I have been able to remind it that it can in fact access the internet using ResearchAgent and it will happily retry the prompt. Just thought I would flag it up and let people know the workaround is to remind it to use ResearchAgent to search with. I haven't seen this behaviour with ChatGPT or Mistral (yet).
Expected: -
Claude models are able to understand they have been augmented with RAG for internet access.
Observed: -
Occasional refusals or contradictory information about web access.
