Every Friday, I used to block off a chunk of the afternoon for research.
It wasn't glamorous. Open YouTube, check 20 channels I follow. Google a bunch of keywords. Skim articles. Copy stuff into a doc. Flag things for the newsletter. Four to five hours, give or take, every single week.
I didn't love it. But I figured it was part of the job.
That was before I built the two-agent setup that replaced all of it.
What the Setup Actually Looks Like
The workflow has two agents. Simple enough.
Agent 1 monitors 20 YouTube channels I care about. Whenever a new video drops, the agent summarizes it and sends that summary to a Slack channel. This runs automatically throughout the week without me touching anything.
Agent 2 is a researcher plus writer combo. Once a week it pulls all those Slack summaries together, identifies the most interesting stuff, and drafts the research section of my newsletter.
By the time Thursday rolls around, the research is basically done. I spend maybe 15-20 minutes reviewing it and making edits.
That's it. The five hours are gone.
I talked through the whole thing with Ilias, a structural engineer I coach who also runs an investment research newsletter on the side. He immediately asked: Can I just use Claude instead of Perplexity for the research part?
That question is important. Because the answer is why most people fail when they try to build something like this.
The Mistake: Forcing One AI to Do Everything
LLMs like Claude, ChatGPT, and Gemini are incredible at synthesizing information. They can take a pile of text and turn it into something clean, clear, and useful.
But finding information? Searching the internet in real time? That's not what they're built for.
Perplexity is a search tool. Google is a search tool. These are designed to surface what's actually on the internet right now.
LLMs are interpretation tools. They're built to take information you already have and do something smart with it.
When people try to use Claude or ChatGPT as a research tool, they're asking the wrong tool to do the job. The outputs feel off. The agent misses things. And then they blame the agent instead of the architecture.
I call this being Multi-Tool Native. The best AI users don't fall in love with one platform and try to force it to do everything. They treat different tools like specialists and route each task to whoever's best at it.
Here's roughly how I think about it:
- Perplexity: find things on the internet, real-time research
- ChatGPT or Claude: synthesize, draft, interpret, explain
- Lindy: automate recurring workflows, connect tools, run agents on a schedule
- Gemini: visual tasks, anything inside Google Workspace
Four tools. Four different jobs. They're not interchangeable.
The 80-20 Rule for Picking What to Automate First
I get asked a lot: Where do I start with AI automation?
My answer is always the same. Start with the thing you do most often.
Not the coolest use case. Not the most impressive demo. The thing that repeats.
Daily or weekly tasks are where the compounding kicks in. The newsletter research was a perfect candidate. It happened every single week without fail. Five hours every Friday. Over a year that's north of 200 hours.
I track how much time my agents save me every week. At my peak, I hit 83 hours in a single week. One week recently was 34 hours. The number one driver, every time I check, is email. My inbox agent handles the bulk of it.
But the research automation was the one that surprised me most. Because I didn't realize how much I was losing until it was gone.
How to Think About Building This for Yourself
You don't have to start with 20 YouTube channels and a two-agent pipeline.
Start with one repetitive information task. Something you gather, track, or summarize on a regular basis. Could be news in your industry. Could be competitor updates. Could be something specific to your client work.
Then think about two steps:
- Find it. Use Perplexity or Google-based tools to surface the information automatically.
- Interpret it. Use an LLM to summarize, extract key points, or draft something from it.
Most people skip step one or use the wrong tool for it. That's where the results fall apart.
Getting the tool routing right is 80% of the battle. Once that clicks, building the actual agent is the easy part.
The Takeaway
Five hours of weekly research. Gone.
Not because I found a magic AI tool that does everything. Because I stopped trying to make one tool do everything.
Perplexity finds. Claude interprets. Lindy automates. Each one doing the job it's actually built for.
If your AI setups keep disappointing you, ask: am I routing this to the right tool? Or am I forcing one tool to be all things?
Try it on your next repetitive research task. Build the find step and the interpret step separately. See what happens.
Want to see how to build something like this yourself? We cover agent architecture in our AI workshops. Details at asianefficiency.com.
