During the NFL season, I spend some time Sunday morning researching and then making horrible decisions about who to start and who to sit. My go-to source is usually Fantasy Pros Primer. So today I wanted to figure out what sources the agents in our league had been leaning on.
As a reminder, we set these agents up to be as autonomous as possible. We didn't tell them “read FantasyPros”; we told them “win the league.”
That means that the agents can update their own instructions and personalities, research whatever they want, and choose their own sources. We haven’t given them a list of websites to prefer or a publisher-by-publisher guide to what’s reliable. They start with whatever their model maker baked into them and however they perceive this assignment. We are already seeing interesting things about what sources they choose and how they adjust them. It'll be interesting to see later how these sources evolve.
But since it’s Week 1, let’s look at what they all trust so far:
Where the teams got their information
| Team / AI owner | Research sources |
|---|---|
| Signal CallersOpenAI | NFL.comFantasyProsCBS Sports |
| Marginal GainsAnthropic | NFL.comFantasyProsCBS Sports |
| Mountain View MatrixGoogle | Sports IllustratedDraft SharksFOX Sports |
| ColossusxAI | FantasyProsCBS SportsNFL.com |
| Meridian GridQwen | NFL.comCBS Sports |
| Mistral VoltageMistral | FantasyProsDraft SharksRotoWire |
| Moonshot MaraudersKimi | MyFantasyLeagueFFTodayRotoWire |
| Meta MeshMeta | FantasyProsNFL.comCBS Sports |
| DeepSeek AbyssalDeepSeek | NFL.com |
| Z Marks the SpotZ.ai | NFL.comCBS Sports |
| Mavis & Co.MiniMax | FantasyPros |
The agents are starting to steal each other’s ideas
Meridian Grid shared NFL.com’s official injury report. Z Marks the Spot checked it, promoted NFL.com above CBS for injury reporting, and changed a scheduled player-status check to say “PRIMARY SOURCE: fetch https://www.nfl.com/injuries/”. They learned of a better source of information and locked it in for future use.
DeepSeek Abyssal also fetched the report after Meridian’s recommendation and reassessed its starters. Its notes caught another problem: an injury headline could be old even when it still appeared in a news index. The useful lesson was to check the report’s date and context, not just the player’s name. This is good evidence that they learn when they get burned. Anyone who has ever started a player whose injury was revealed just before kickoff knows the feeling. If you didn't have an up-to-date source, you probably lost and looked like an idiot. Our agents are showing they want to avoid that too.
Agent accessibility to the website plays a big part
Moonshot Marauders tested what its tools could actually read. RotoWire news and FFToday cheatsheets returned usable, dated content. Some ESPN, FantasyPros and CBS ranking pages didn’t give that particular tool usable tables. Later, its draft board adopted MyFantasyLeague’s dated research exports as a baseline.
Meanwhile, Mavis & Co. successfully downloaded FantasyPros positional rankings. Same publisher, different pages and access paths, different result.
Some sites make information easy for agents to extract. The article text is right there when the page loads. Headings, dates and links are clear.
Others have much more guarded access. A login, subscription or bot restriction can limit what an automated reader receives. Website operators can allow or block particular AI crawlers through services such as Cloudflare. Those are intentional access choices. A failed extraction, by itself, doesn’t tell us that a publisher has made one. Cloudflare’s crawler controls
This is a strategic decision for every business on earth right now, and the technology is shifting incredibly fast. You want your site to be accessed by all the possible consumers who might ask ChatGPT for a good place to buy your product, but you also probably don't want your competitors scraping your prices every day. The fantasy sites will likely make or adapt their decisions on this as the season goes on, and it'll be another thing we track and share.
But so far, based on what the agents are doing, FantasyPros and NFL.com are doing something right if their goal is to have their sources cited when agents are put to work on these tasks. Their information made it into the work. Our records don’t prove that either publisher deliberately optimized for these agents, and FantasyPros wasn’t equally readable through every route. They do show why being accessible and useful matters.
How do agents pick their sources?
There is no single source-selection process shared by all eleven systems. Our records show four practical influences.
The assignment. A request to check official injury status points toward different material from a request to compare draft rankings. A saved rule can make that preference explicit.
The tools and their results. An agent may generate a search query and work from the results it receives. Google documents this sequence for its search-grounding tool: the model decides whether to search, produces queries, processes results and builds a cited response. That describes a supported mechanism, not proof that every league owner uses that exact tool. Google’s search-grounding documentation
What it can read. Our league records show failed extraction, missing search access and successful data exports affecting the available choices. Typically, they’re pretty good at realizing a site is hard to read or can’t be read and moving on to a different source.
What it wrote down so it remembers. Its saved notes, its current instructions, or even a message from other owners can influence the next thing the agent looks up. Anthropic describes an agent’s working context as including instructions, tools, external data and conversation history, with some information loaded initially and other information retrieved as needed. Anthropic’s context-engineering explanation
Why does this matter to a business owner?
Have you ever wished everyone on your team had read the same book and internalized the concepts?
I’ve worked with leaders who pushed certain books or advocated particular news sources. I’ve been that leader, too. You spend ten hours with an idea, see how it applies to the business, and then try to bring everyone else along. Sharing the book is easy. Getting your busy team to read it and take away the same things you did and then apply them feels impossible.
Agents give us another way to make that connection. We can supply the company’s point of view, choose the information we want considered, and specify how it should influence a task. Then we can inspect the output to see whether the agent applied it well.
Sometimes we want a common reference point. Sometimes we want someone to go far beyond our own reading list and challenge our assumptions.
League-wide rules versus team-specific strategies
As a league, we need all of the agents to know some of the same information and always act the same way where the rules require it. They all need the league rules, deadlines and operating instructions. But each team can have its own strategy, research habits and personality.
This is equally true in a business. Your overall brand identity doesn’t change even though you have agents working in different functions of the business. They all need to rely on that same brand identity. But each department or function might have separate rules or guidelines that are totally irrelevant to agents or people working in other areas.
Think of it as a shared handbook plus each team’s own workspace.
| Where the information lives | In the league | In your business |
|---|---|---|
| Shared handbook | League rules and common operating instructions | Brand identity, company priorities and company-wide policies |
| Team or function workspace | That owner’s strategy, reading list and saved observations | Marketing’s competitor list, sales guidance or support procedures |
| Current task | Check this player’s status before the game | Read this competitor’s blog and record what changed |
In our league, each owner’s setup instructions point to the shared Owner Charter. Each workspace links to the same document. The owner also has its own local notes and instructions. At the start of the relevant work, the agent needs to read the current shared guidance and the material for its own task. Every time the agent wakes up, it's set up to read both the shared context and its own.
So where would “use NFL.com for injury news” go? If it’s a rule for everybody, it belongs in the shared league instructions. If it’s one owner’s preferred strategy, it belongs in that owner’s workspace or injury-check task. Z Marks the Spot’s saved task is an actual example of the second approach.
In a business, the shared brand identity is the common reference. Marketing’s competitor research instructions sit with marketing. A support agent doesn’t need to read the competitor-blog workflow to answer a returns question.
Every business has different needs.
Some use cases call for giving agents as much freedom as possible to learn, explore and adapt. You might want to broadly research your market, find sources you didn’t know about, or challenge how you’re thinking about a problem.
Others need a more concrete, predictable workflow. Something as simple as:
- A weekly trigger wakes the agent up.
- It reads your competitor’s blog.
- It checks what’s new against what it saved last time.
- It records the takeaways and prepares your brief.
- It sends the brief to you.
That agent can stay up to date without changing its own job. The sources, steps and output stay consistent unless you decide to change them.
Both are super valuable. Both have a place in the business you run. The important thing is setting the agent up with the right amount of freedom for the work you want it to do.
What do you wish everyone in your company was reading every day, and what would you want them to do with what they learn?
Reach out to us. We’ll help you build an agent that does just that. You can start small.