My research agent cross-references sources, flags contradictions, and delivers briefings. Here's how I built it and what it actually does.
Every Monday morning, a research briefing lands in my inbox. It covers industry developments from the past week, competitor movements, regulatory changes in the markets I operate in, and emerging trends that might affect my businesses.
This briefing used to take me five hours to produce manually. Scanning RSS feeds, reading articles, cross-referencing sources, checking if the “breaking news” from one outlet was actually confirmed by another.
Now a research agent does it. The briefing is in my review queue by 7 AM. I spend twenty minutes reading it. The quality is better than what I produced manually because the agent checks more sources and never skips a step because it is tired on a Friday afternoon.
The agent runs three types of research on different schedules.
Weekly industry scan. Every Sunday night, the agent scans a curated list of sources for developments in AI, technology consulting, and the North African tech market. It produces a summary organized by topic, with links to the original sources and a relevance score for each item.
The relevance scoring is the key differentiator from a simple RSS aggregator. The agent knows what I care about (agent architecture, AI consulting business models, regulatory changes in Algeria and the UK) and scores each item against those interests. Items below a threshold are listed in an appendix but not included in the main briefing.
Ad-hoc deep dives. When I need to understand a specific topic, I hand the agent a research brief. “I need to understand how Company X’s new agent framework compares to our architecture.” The agent searches public sources, reads documentation, analyzes technical blog posts, and produces a structured comparison.
These deep dives used to take me two to three hours of reading and note-taking. The agent produces a first draft in about fifteen minutes. I spend another fifteen minutes reviewing and asking follow-up questions. Total time: thirty minutes instead of two to three hours.
Competitive monitoring. A subset of the research capability focused on specific competitors and market players. The agent tracks public announcements, job postings (which signal strategic direction), content output, and pricing changes. A weekly summary flags anything that looks like a strategic shift.
This is not espionage. Everything the agent reads is publicly available. The value is not access to information. The value is consistent, systematic attention to information I would not have time to track manually.
The exact architecture, memory layers, and delegation patterns I use to run 50 agents across two businesses.
Get the AI Agent Blueprint →The research agent is one of my specialized agents. It connects to several MCP tool servers.
Web search and browsing. The agent can search the web and read web pages. For the weekly scan, it works from a seed list of sources (industry blogs, news outlets, company blogs) and follows relevant links from there.
Memory search. Before writing any research output, the agent searches my internal memory layers. This lets it reference past research, connect new findings to previous analysis, and avoid repeating information I already know.
File operations. Research outputs are saved as structured documents with metadata: date, topic, sources, confidence levels. These accumulate into a searchable research archive.
Intelligence pipeline. The research agent feeds into and draws from a broader intelligence pipeline that aggregates signals from multiple sources. The weekly briefing is the most visible output, but the pipeline also fires alerts for high-priority items that should not wait for the weekly summary.
The part of the research agent that took the longest to get right was source evaluation.
Not all sources are equal. A technical blog post from someone who built the system is more reliable than a journalist’s summary of an announcement. A peer-reviewed paper carries more weight than a conference talk abstract. A company’s SEC filing is more reliable than their marketing page.
The agent applies a source hierarchy when evaluating claims. Primary sources (official documentation, filings, direct announcements) outrank secondary sources (news coverage, blog summaries). When a claim appears only in secondary sources without a traceable primary source, the agent flags it as “unverified” rather than presenting it as fact.
Contradiction detection was the other hard problem. When Source A says a company’s revenue grew 30% and Source B says 15%, the agent does not pick one. It presents both with their sources and notes the discrepancy. I decide which to trust based on the source quality and my knowledge of the subject.
This matters more than it sounds. Early versions of the research agent confidently reported incorrect numbers because it averaged contradictory sources or defaulted to the most recent one. Now contradictions are explicitly surfaced, and I make the judgment call.
Five hours per week is 260 hours per year. At the level of research quality the agent produces, those hours were not fun or creative. They were scanning, reading, noting, cross-referencing, and formatting.
Here is what I do with those five hours now.
Two hours go to client work. Additional time on client delivery means better outcomes. This directly affects revenue and reputation.
One hour goes to strategic thinking. I read the agent’s briefing and think about implications. What does this trend mean for my positioning? Should I adjust my service offerings? Is a competitor’s move something I should respond to or ignore?
Before the research agent, the strategic thinking was squeezed into whatever time was left after the research itself. Now the research is done and the thinking gets dedicated time.
Two hours go to content. The research briefing is a source of content ideas. Topics that I found interesting, developments I have an opinion about, trends that connect to my experience. Some of my best LinkedIn posts and blog articles originated from items in the weekly briefing that I wanted to comment on.
If you want to build a similar capability, here is where to start.
Define your sources first. Do not point the agent at “the internet.” Give it a curated list of 20 to 30 sources that consistently produce relevant, reliable content. Expand the list over time as you discover new sources.
Define your interests explicitly. The agent needs to know what you care about to score relevance. Write it down in detail. “AI” is too broad. “Agent architecture patterns, multi-model routing strategies, and AI consulting business models” is specific enough to produce useful filtering.
Build contradiction detection early. If the agent reports conflicting information as settled fact, you will lose trust in the output quickly. Surfacing contradictions is more valuable than resolving them automatically, because the resolution often requires judgment the agent does not have.
Review everything at first. For the first month, review every output carefully. Note what the agent gets right, what it misses, and what it includes that is irrelevant. Use these observations to refine the source list, relevance scoring, and output format.
Expect the first version to be mediocre. My research agent’s first output was a wall of text with no structure, no source attribution, and no relevance filtering. It took three weeks of refinement to reach the current quality. The investment pays back quickly once the system is tuned.
The blueprint includes the research agent configuration, source list template, and relevance scoring framework.
Research is the easiest knowledge work to automate well. The inputs are public. The process is systematic. The output format is structured. If you are spending hours each week staying informed about your industry, an agent can do 80% of that work at higher consistency than you manage on a busy week.
Subscribe to The Signal
One issue per week. The biggest AI build of the week, the top five videos worth your time, and an operator's read on what actually matters. For builders who ship, not dreamers who scroll.
No spam. Unsubscribe anytime. Just The Signal.