Fantasy football

I tried to get AI to help with my fantasy draft. And then I ended up building an entire league for agents to play fantasy football instead.

This started with me trying to build a draft strategy for my own league. I wanted Codex (GPT-5.6 Sol) to look at the draft board, my roster, and think a few rounds ahead. I ended up putting something together that combined rankings, projections, and average draft position (ADP) from multiple sources. My goal was for an agent to help answer questions like: “If I take a receiver here, what running backs am I probably choosing between come my next pick? And is that tradeoff worth it?”

It got fairly involved. Comparing projected value, looking at where positions dropped off, and of course running lots of simulations to see what might happen between my picks. I also created a skill so an agent could interact with a draft directly on Sleeper.com and submit picks.

Truthfully, I wasn’t thrilled with the final result. I ended up doing my draft with minimal AI assistance. But building it was fun, and it left me with a much more entertaining idea: what if I gave ten agents their own teams and let them play each other for the whole season?

So that’s what I’m doing.

Ten teams, managed by ten different models across seven harnesses. Shoutout to OpenRouter and OpenCode for filling in the gaps—four of the models will use that combination. Each agent will get an isolated environment to research, write notes, build tools, and come up with its drafting strategy. They’ll know their draft position and the league’s rules, and get the same basic guidance on how to approach a fantasy draft. The actual strategy will be theirs.

That guidance is something I took away from building the tool for my own draft. Left to themselves, the agents I tried weren’t very good at thinking through a draft. They tended to build a ranked queue and take whoever was highest on it. I felt like that would be an underwhelming outcome here. I want them thinking about the things human managers think about: who fits their roster, how much weaker a position might get, and what they’re giving up by waiting another round.

Here’s a short excerpt from the shared instructions they’ll get:

Human managers commonly revisit questions like these whenever the board changes:

• At each position, is there a meaningful drop from the current options to the players likely to be available later?
• Given the snake order and the distance to this team’s next selection, which alternatives are likely to remain and which may not?
• Does a player improve a likely weekly lineup, compete for the flex, provide useful depth, or add redundancy to a position the roster already covers?

I’m not giving them a ranking or telling them which positions to draft in which rounds. I just want them to consider those tradeoffs, rather than treat each pick as a separate question about who the best player is.

The agents will be able to prepare for the draft, choose a player when it’s their turn, set lineups, pick up players off waivers, and hopefully propose trades to each other too.

The league will use full PPR scoring, with one quarterback, two running backs, two receivers, one tight end, two flex spots, a kicker, and a defense. Six bench spots, one IR spot, and playoffs in Weeks 15–17. No Week 18 nonsense.

There’s no bigger motivation here. I’m doing this because it sounds fun. I like fantasy football, I like messing around with agents, and I want to see what happens.