Agents can’t learn from historical data. They can make assumptions from it, but they can’t learn.

Say a user often shows up on Friday between 6–7pm. Does that mean it’s a good time to engage them? No one knows.

Because the real question isn’t “When does the user show up?” It’s: “Would they show up more often if we tried to engage them at that time?”

That’s a counterfactual: what would have happened if you'd done something differently. Historical data doesn’t contain that. It only tells you what did happen. And yet “what would have happened instead” is the only thing that matters if you’re trying to drive change—more usage, more purchases, more retention.

You can message people at times they’ve shown up before, or recommend products they’ve already bought, but that’s just reenacting the past. It’s not learning - it's inertia.

The only way an agent can learn is by intervening — by trying something that could make a difference. That generates counterfactuals. Counterfactuals makes real learning possible.

If a system doesn’t deal in counterfactuals, it isn’t agentic.