Agents for Humans: Why CAST Replays the Same Play Under Different Rules
Generating a new mechanic is only half the problem. CAST records a real playtest and replays the same input from the same initial state under old and new rules, so designers can compare the mechanic instead of trying to recreate the moment.
Generation was not the interesting part
When we started building CAST for Agents for Humans, the obvious feature was natural-language mechanic editing.
A designer could say:
“Only pick up the cargo when I approach slowly.”
A Strands Agent could interpret the request and create a new rule.
That worked as an agent task.
But it did not yet solve the most frustrating part of playtesting.
The designer still had to answer:
Did this new rule actually improve the interaction I just experienced?
And to answer that, they usually have to play the situation again.
The second attempt is never exactly the same as the first.
The cursor path changes.
The timing changes.
The angle changes.
The speed changes.
So we shifted the center of CAST from generation to comparison.
The product became:
Play once. Change the rule. Replay the same play under both versions.
We call this counterfactual playtesting.
Preserve the moment that caused the idea
Imagine a designer is moving a player past a cargo object.
The cargo attaches too easily.
That specific interaction is what causes the designer to think:
“This should only attach when I approach slowly.”
Normally, the moment disappears as soon as the designer stops playing.
They make the change and then try to reproduce what happened.
CAST instead records the playtest itself.
For each play, we preserve data such as:
1
2
3
4
5
initial world state
ordered pointer input
relative timestamps
movement values
current mechanic versionEnjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article