How Can I Improve the Source of My News?
Early on, the research agents used general web search for news gathering. It worked, but web search returns a mix of everything — ads, SEO-optimized content, outdated articles, irrelevant results. Getting clean, current, high-signal news required filtering through a lot of noise.
So I asked my system: how can we make this better? It came back with the idea of switching to Tavily.
Tavily is a search API built specifically for AI agents — it returns clean, structured results optimized for LLM consumption rather than human browsing. Results are deduplicated, ranked by relevance and recency, and stripped of the junk that makes general web search frustrating for automated pipelines.
The practical difference: the research agents now spend less time filtering and more time analyzing. The signal-to-noise ratio in the overnight briefs improved noticeably after the switch. Less “here’s a listicle from 2023,” more “here’s the actual Reuters article from this morning.”
The Flow End-to-End
Here’s how a typical overnight cycle runs:
- ~10 PM PT — AIResearcher, TechResearcher, and FinanceCrawler activate, each pulling from Tavily and other sources for their domain
- ~11 PM PT — Agents send findings to OvernightAnalyst
- ~11:30 PM PT — OvernightAnalyst synthesizes and sends to StockPicker
- ~5:45 AM PT — StockPicker generates picks and delivers morning brief
- 6:00 AM PT — Brief delivered, 45 minutes before market open
The whole thing runs while I sleep. By the time I’m reading the brief over coffee, the agents have already done hours of work.
What Makes This Different from a News App
A news app shows you what happened. The research swarm shows you what happened and why it might matter to your specific portfolio and interests. The agents accumulate context — they know which sectors I follow, which macro themes I’ve been tracking, which positions I’ve asked about before. The briefings aren’t generic market summaries; they’re filtered and framed for my situation.
That personalization compounds over time. The agents that have been running for two weeks produce better output than they did on day one, not because the model improved, but because their memory improved.
Images, Videos? Is It Possible?
NanoClaw defaults to Claude, which is great — but it doesn’t generate images or videos. Not that I needed that anyway. But it doesn’t hurt to try, right?
After chatting with my home AI, it pointed me to Gemini Veo 3.1 for video and OpenAI DALL-E for image generation. Before I even entered my first prompt — $$ — more API connections and tokens needed, this time for OpenAI.
The results weren’t bad. But the cost was a different story. Veo 3.1 is expensive — a 15-second video cost more than I’m comfortable paying. Images, on the other hand, were more affordable than I expected.