Marketing

How we built a market monitoring system with AI agents

The New Rails cover reading Market Monitoring With AI, with the subtitle MUCH HARDER THAN IT LOOKS below it, next to a glowing blue radar with signals on its sweep

MOIC Digital monitors the crypto and financial markets with Watchtower, a system of AI agents that runs every business day. Code collects news from more than 160 curated sources and 32 public X lists, keeps only the last 24 hours, and an editor agent checks a theme memory before writing. The result is a daily Notion page with the news, the saturated themes, the angle nobody is exploring and a story idea for each audience.

Monitoring a market is more complex than it looks. By default, AI agents struggle to find news in real time, follow what happens on social media and separate content with new ideas from content that only repeats the consensus. Here is how curated sources, managed by the team, fixed that for us.

Every marketing and sales team has run this test: open a chatbot and ask for “the most important crypto and market news of the day”. The answer comes in seconds, well written and organized in bullet points. By the second week, the team notices it is almost the same every day, half the links do not open, some of the news is from last week and nothing in it turns into a better post than any competitor’s.

At MOIC, analysts follow dozens of clients among financial companies, each one speaking to different audiences: professional traders, institutional allocators, DAO treasuries, developers, decentralized finance (DeFi) users, the crypto audience in Latin America. Every day they need two things: to know what happened in the market and to have an angle of their own to turn it into content for each client.

Watchtower is the system of agents MOIC’s AI Squad built for this. It has run every business day since May 2026. Building it, we ran into a series of problems that show up for anyone using AI agents to monitor a market. Almost all of them have the same origin: the agent depends on the web search tool built into it. That search works well for general questions, but it is weak at three things a market radar needs: news from the last few hours, what circulates on social media, and the full, specific content of niche sources. The decision that changed the result the most was to stop depending on it and build a base of curated sources that the team manages and the code collects directly.

A simple prompt (“Give me the news of the day”) Watchtower, a system of agents
Sources Whatever the search engine returns first: big portals and SEO 164 audited sources and 32 X lists, tagged by audience
Date Mixes today with last week A single 24-hour window; with no confirmed date, the item is dropped
Links Sometimes rebuilt from memory and broken Taken from the feed or the API, never from the model’s memory
Memory Starts from zero with every request and repeats themes The Theme Ledger records themes, days in a row and angles used
Reader An average summary for no one in particular Two readings per persona, each with a story idea
Language Jargon and lists of names without what happened Who, what, how much, when; technical terms explained
Angle Only the consensus Consensus flagged, alpha, takes from named authors and creative triggers
Destination Stays in the chat of whoever asked A Notion page that feeds other agents and the team

How does Watchtower work?

Watchtower combines code, AI agents and a persistent memory in a routine that runs on its own every morning.

  • Curated sources. More than 160 active sources (newsletters, central bank research, regulators, technical blogs, onchain data, prediction markets) and 32 public X lists, each tagged with the audiences it serves.
  • Collection in code. Scripts pull the feeds and the X lists in parallel and, before any AI reads them, drop everything older than 24 hours or without a date.
  • Theme memory. Before writing, the agent checks the Theme Ledger, a table that records how many days each theme has appeared and which angle has already been used.
  • Editor agent. It classifies each theme (new, developing or saturated), writes the news following fixed writing rules and builds the sections of the page.
  • The day’s Notion page, with news, opinions from relevant authors, saturated themes with the alternative angle, what circulates on X, creative triggers and a reading for each audience.
  • Derived agents read that page and produce other deliverables: a roundup per client, post suggestions for each client’s X profile, 20 ready-made posts and the link posted to the team’s Discord.

Work is split as in our other systems: what needs to be exact (collection, date filtering, counting) lives in code, what requires reading and judgment (what counts as news, what the angle is) goes to the agent, and the decision to use a story idea with a client stays with the analyst. Versa, our proposal agent, splits its work the same way.

What goes wrong when you just ask for “the news of the day”?

Each part of Watchtower exists to solve a problem we hit when using agents in a simple way. If you are building something similar, this is the list to check.

When you ask an agent for news, it uses the search tool that comes with it. That search has three limits for market monitoring:

  • Real time. The search index takes a while to register what was published in the last few hours, and the results mix today with last week.
  • Social media. Web search barely sees what circulates on X, where much of the market conversation happens first.
  • Rich content. The agent receives titles and short snippets. The full text of a newsletter, the numbers in a report or the argument of a technical post rarely arrive whole.

The result is that the agent reads what shows up first: big portals, content optimized for search and the same headlines everyone has seen. What sets an analysis apart (a researcher’s technical blog, a central bank note, a protocol’s governance forum) is almost never at the top.

How we solved it: we took web search off the main path. Watchtower collects directly from sources the team chose and manages: the publications’ feeds and pages, data APIs and the X lists through the platform’s API (X API documentation). Web search stayed only for the few paywalled sources, where it is the only way to read the summary. To manage these sources, we keep an official list, audited to know whether an agent can access each one and how. Each source is tagged by access method: RSS (a structured feed), HTML (a simple page), API or PREVIEW (title and summary only, because of a paywall). Sources that require login, block robots or have died go to an archived list and are never queried. In June we added 45 “alpha sources”, such as primary onchain data, public positions filed with regulators, prediction markets and regional research from Asia and Africa.

Access method Active sources What it means
RSS 71 A structured feed, the cheapest
HTML 40 A simple page, without JavaScript
API 29 Onchain data, prediction markets
PREVIEW 24 Title and summary only (paywall)

Watchtower’s 164 active sources by access method, plus 32 X lists (12 daily, 20 in rotation).

One curation decision made a difference: each audience gets sources on its main subject and also on adjacent interests. A trader reads about trading, but also about macroeconomics, behavior and AI. The adjacent sources are where the connections that become original content show up.

2. Old news passes as today’s news

Website home pages mix pinned articles, old content and whatever is trending. An agent that reads “the first items on the page” delivers week-old news with the confidence of someone talking about today. This was the first problem our analysts pointed out.

How we solved it: the first thing the system does is calculate a single 24-hour window, which applies to every script and agent. Items outside the window are dropped at collection. When an item’s date cannot be confirmed, the rule is to drop it. RSS makes this easy, because every item in a feed can carry its publication date and its link (RSS 2.0 specification). We lose a little coverage and gain the confidence that everything on the page is from that day. The agent is also not allowed to use expressions like “recently” or “in recent weeks”, which usually give away an old item.

Language models sometimes shorten or rebuild URLs from memory, and the link looks right but does not open. We had this problem in June, with links to blogs that did not exist.

How we solved it: the link for each news item comes from the feed itself or from the API response, never from the model’s memory. Each news item has a keyword linked to the primary source, and the X post quoted comes with the exact address returned by the API.

4. Without memory, the agent repeats the same theme every day

A chatbot starts from zero with every request. If the market spends three weeks talking about interest rates, you get the same headline about interest rates for three weeks. The reader stops reading, and the content that comes out of it sounds like everyone else’s.

How we solved it: the Theme Ledger, a Notion page the agent reads at the start and rewrites at the end of each run. For each theme, it stores how many days it has appeared, its current state and the angles already used. Before writing, the agent classifies each theme:

State When In the digest
Emerging Has not appeared in the last 5 days Goes into the news normally
Developing Appeared on 1 or 2 recent days and has a new fact Goes in with only what changed, in one line
Saturated Saturation score of 6 or more, or 3 days in a row Leaves the news and goes to Saturation & Alpha
Cooling Lost volume or disappeared from X At most a short note, or nothing
Archived Back to zero Leaves the radar

The main rule is simple: a theme the reader saw as news yesterday does not come back as news today. Either it has a new fact, and goes in as a short line with what changed, or it has saturated, and leaves the news section to get a different treatment.

5. Saturated does not mean irrelevant

When a theme saturates, its value moves from the fact to the angle. Everyone already knows bitcoin went up. Few know what that means that is different.

How we solved it: the 🌀 Saturation & Alpha section. For each saturated theme, the agent scores saturation from 0 to 10 (days in a row in the digest, how many X lists talk about it, engagement and coverage on the big portals) and delivers four parts:

  • Consensus: what everyone is already saying, explicitly marked “do not post this as is”.
  • Alpha: the angle off the curve, looked for in four places: specialized profiles with little engagement, people who disagree with the consensus with an argument, the theme’s effect on another market, and what nobody is discussing.
  • Immediate implications: what should happen in the next few days, without speculation.
  • Second- and third-order implications: creative scenarios over months to years, marked “(speculative)”.

Real excerpt from the Watchtower Daily of September 4, 2026: 🪙 Bitcoin, ETFs and a four-month high

📛 Consensus, do not post this as is. Bitcoin ETFs took in US$731 million, the biggest day since January, and BTC went back above US$80,000. The ready-made reading is “institutions are back in force”. Saturation 9/10: 54th day on the radar, on the front page of every crypto portal.

🎯 Alpha. The four-month high comes with the network almost idle: bitcoin’s fastest fee is at 4 satoshis per byte, the floor of a network with no queue. The price rises on paper (ETFs and derivatives) while almost nobody actually moves bitcoin. Nobody is pointing out the mismatch between a record price and an empty chain.

Suggested story: “Bitcoin hit a four-month high with the network almost idle at 4 satoshis per byte: this is a paper rally of ETFs, not of people using bitcoin.”

The angle used is recorded in the Theme Ledger, so the agent picks a different one the next day. In October, the long-term interest rates theme reached its 30th day in a row on the radar, with a different angle every day. We have written about why markets reward the angle over the headline in how narratives drive attention and market cycles.

6. A generic answer, with no defined reader

“The news of the day” does not say for whom. The result is an average summary that does not serve any audience well.

How we solved it: on every run, Watchtower reads MOIC’s persona base in Notion (today there are 8, from Pro Trader to Web3 Marketer). For each one, the agent writes two readings of the day:

  • Direct view: a thesis you can defend in a client meeting, connecting two or three news items, with a “why it matters” line and a suggested story.
  • Unconventional view: a provocative reading that connects real signals in an unexpected way, also with a suggested story.

Because the persona list comes from Notion, adding a new audience does not require changing the agent. That is how the Web3 Marketer persona came in in June, with its own sources.

7. Jargon and a “menu of topics”

Agents tend to write headlines that list names without saying what happened (“five AI models launched in a row, with a new benchmark redefining what counts as the frontier”). For anyone outside the field, that says nothing.

How we solved it: writing rules in the skill itself, with examples of what is good and what is bad. Each news item has to say who did what, how much and when, use a concrete verb (“cut”, “launched”, “bought”) and explain in parentheses every technical term the first time it appears. The test: would a marketing intern, a designer and an engineer understand the sentence without opening Google?

Example from the skill’s writing rules, before and after

Before: 🪙 BTC hit a new high: market in risk-on mode with institutions buying.

After: 🪙 BTC passed US$145,000 for the first time yesterday after BlackRock’s spot ETF (an exchange-traded fund that holds BTC directly) recorded US$2.1 billion of inflows in the week, its biggest intake since approval in January 2024.

8. Only facts, no creative raw material

Even a perfect list of news does not help much if you need to write a different post. It lacks a hook, a comparison and an opinion.

How we solved it: two sections that exist only for this. Takes of the day gathers opinions and predictions outside the consensus from authors named in the collection, such as fund managers, researchers and newsletter writers, with one line explaining why the take is interesting. Side layer is a deck of the day’s creative triggers: anniversaries, quotes from thinkers, a counterintuitive statistic, a paradox, an unlikely analogy, an empirical law applied, a naive question, an alternative headline. Each persona’s unconventional view has to use at least one of these triggers.

Section What it holds Type
News of the day Up to 25 facts from the last 24 hours, one sentence each, with a link Fact
Takes of the day Off-consensus opinions from named authors Opinion
Saturation & Alpha Consensus, alpha and implications of saturated themes Angle
On X today What circulates on the lists, with quoted posts Signal
Side layer Anniversaries, quotes and 10 creative triggers Raw material
Narratives by persona Direct and unconventional view for 8 audiences, with a story idea Action
Sources consulted Where each item came from Trail
Source health What failed today and alerts for repeated failures Control

The anatomy of Watchtower’s daily page.

9. The answer dies in the chat

A chatbot answer stays in the conversation of whoever asked. Nobody else on the team sees it, and no other process uses it.

How we solved it: Watchtower publishes a Notion page, in a stable format that other agents can read. Derived agents run on top of it:

  • Watchtower per client: crosses the news with each client’s profile (sector, narrative, competitors, keywords) and writes a block ready to send plus an internal note for the analyst.
  • Comms Channels: reads the recent posts on each client’s X profile to understand the tone and suggests what would make a post in that tone.
  • 20 Viral Posts: turns the day’s page into 20 posts in English, using a page of examples of good content as a reference.
  • Discord posting: posts the day’s links in the team’s channel.

Watchtower’s content also started feeding client deliverables, such as a weekly news thread and LinkedIn posts.

10. Silent failure

When a chatbot cannot access a source, it simply does not mention it, and you never find out.

How we solved it: every page ends with a source health section: how many feeds responded, which failed and why (an anti-robot block, an address that changed, an expired API key). A source that fails three days in a row raises an alert, and five days in a row produce a proposal to archive it. This section is how we saw, in September, that the X API key had been expired for 18 business days.

How to build your own

The steps below apply to any market radar built with agents, even if your subject is a different one.

1. Define the readers before the sources. Write the personas who will consume the radar and the adjacent interests of each one. Keep this somewhere the agent reads on every run, so you can change it without touching the skill.

2. Curate and audit the sources. For each source, record how an agent accesses it (RSS, HTML, API or summary only) and which personas it serves. Sources that require login or block robots go to an archived list. Prefer RSS: a single file brings dozens of items with a date and a link, for a fraction of the cost of searching the web.

3. Collect in code. Scripts collect in parallel, filter by the time window and cap the number of items per source (we use 5) so noisy sources do not dominate. The agent only receives what has already passed the filter.

4. Give the agent memory. A simple table with theme, days in a row, state and angles used solves repetition. The agent reads it at the start and updates it at the end of each run.

5. Write the writing rules into the skill. Include examples of good and bad. Abstract rules (“be clear”) do not change the result much. Before-and-after pairs do.

6. Separate fact, opinion and angle. News carries only facts. Opinions go to their own section, with an author. Your angle (alpha, implications, story ideas) goes to another. That way the reader knows what they are reading in each part.

7. Publish in a format other agents can read. A stable page in Notion (or anywhere with an API) becomes input for derived agents. That is where a radar turns into a content production line.

8. Monitor source health. Record failures every day and create alerts for repeated failures. Without this, quality drops slowly and nobody notices.

What is still open

Some lessons came from problems we are still solving:

  • External dependencies stop everything at once. When the X API key expired, the X section stayed empty for weeks in the automatic runs, because the alternative (reading X through the browser) requires someone logged in.
  • Running on one person’s computer is fragile. Discord posting stopped for about two weeks, and the hypothesis is that it depended on one specific machine being on.
  • The next step is distribution. We are testing a “market update” sent to clients and content on MOIC’s channels using Watchtower as the source, as well as reusing the story ideas in formats that AI search engines read.

Principles

  • Code to collect and filter, the agent to read and write, a person to decide what goes to the client.
  • When in doubt about the date, drop it.
  • The link comes from the source, never from the model’s memory.
  • Memory between runs is what separates a radar from a summary.
  • A saturated theme changes section: it leaves the news and gets a new angle.
  • Every output is input for the next agent.

Want to apply this to your marketing?

We help financial companies selling to businesses use market intelligence for content that earns attention. Our services explain the work, and you can talk to us.

FAQ

What is Watchtower?

Watchtower is MOIC Digital’s market monitoring system, a set of AI agents that follows the crypto and financial markets. Every business day it publishes a Notion page with the last 24 hours of news, saturated themes with a new angle and a story idea for each audience.

Why not just ask a chatbot for the news of the day?

A chatbot depends on its built-in web search, which is weak at news from the last few hours, social media and the full text of niche sources. It also has no memory between requests, so it repeats themes and sometimes rebuilds links that do not open.

How do you keep old news out of an AI news digest?

Watchtower applies a single 24-hour window at collection, in code, before any AI reads the items. When an item’s date cannot be confirmed, it is dropped.

How does the agent avoid repeating the same theme every day?

It reads and updates the Theme Ledger, a table with each theme’s days in a row, state and angles used. A saturated theme leaves the news section and gets a new angle in Saturation & Alpha.

Written byMOICMOIC Digital

Let’s turn these insights into action and revenue.

We work with B2B onchain finance teams on positioning, demand and pipeline.

Book a call

Keep reading

All posts →