AWS Builder Center

Agents for Humans: Designing a Safe Mechanic Patch Loop with Strands Agents

How CAST uses Strands Agents, constrained tools, versioned mechanics, and explicit failure handling to turn natural-language design intent into executable game-rule patches without letting the model freely rewrite the application.

Why we did not let the model write arbitrary game code

CAST is a live playtest agent for game and interaction designers.
A designer can play a prototype, describe a mechanic change, and compare the exact same recorded play under the old and new rules.
The natural-language part sounds simple:
“Only attach the cargo when I approach slowly.”
But implementing that safely raised a more interesting question:
How much control should the agent actually have?
One option would be to give a model access to the source code and ask it to rewrite whatever is necessary.
We deliberately did not build CAST that way.
Instead, CAST uses the Strands Agents SDK with a constrained mechanic runtime and explicit tools.
The agent can reason about what should change, but the actual changes have to fit a bounded set of executable primitives.
That makes the system easier to inspect, replay, patch, and revert.

Intent is flexible; execution is constrained

The designer speaks in mechanic-level language.
For example:
“Only attach when I approach slowly.”
The agent does not need the designer to say:
1
2
3
4
distance < 0.30
AND
speed < 0.25
→ attach
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article