for research agents, do you let a deep research api run the whole loop or build your own on top of search + extract?

been reading up on deep research apis for agents and there are basically two camps one call does everything. you send a prompt and get back a cited report (gemini deep research, openai, octen, tavily research etc) you build the loop yourself. you only use search/scrape primitives (firecrawl, exa, tavily search) and your own model does the planning and synthesis option 1 looks way easier but you can't really see why it picked the sources it picked, and some of these runs take 45-60 mins. option 2 gives you control but now you're the one dealing with anti-bot pages, retries and making sure 30 parallel queries don't fall apart also the pricing is all over the place. one is $0.20-3 per run, another is tokens + tool calls, another is credits that can be anywhere from 4 to 250 per request. how are you even estimating the cost per finished task before you ship? for people actually running this in prod: * which camp are you in and why * has a one call api ever beaten your own loop on quality? * what broke first when you scaled up just curious, i'm trying to decide before i build something