Creator PSA: Watch Your Mushie Usage With Agent Coding
Oncin ·
I just learned this the expensive way 😭
If you use agent models to edit larger custom worlds or UI files, watch how many generations and tool steps they use.
The expensive part may not be the feature itself. A small UI task can become surprisingly costly when the agent repeatedly carries a large project context through separate reasoning, reading, editing, and validation cycles.
Things creators should watch for:
repeated planning or re-planning
stalled generations that make no progress
many tiny sequential edits instead of a batched patch
repeatedly re-reading the same large files
validating after every small change
investigating unrelated systems
continuing to polish after the requested feature already works
Large worlds or large single-file UIs can amplify this because every additional agent generation may carry a substantial context baseline again.
Cost constraint I’m using from now on
Inspect only code directly relevant to the task.
Search first, then use narrow targeted reads.
Do not re-read already inspected regions unless genuinely necessary.
Decide the complete patch before editing.
Batch independent edits into the same write pass where possible.
Do not spend separate generations on tiny CSS/function/wiring changes.
Validate once after implementation.
Fix only concrete validation failures.
No unrelated cleanup, refactoring, optimization, or extra review.
Do not repeatedly reconsider decisions already specified.
Once the requested behavior works and validates, stop.
For straightforward UI patches, it may also be worth using cheaper/faster models and reserving expensive reasoning models for difficult debugging, architecture, or cross-system changes.
Exact token accounting depends on the platform and model, but the main thing to watch is agent iteration count.
A lot of high-context generations can burn through Mushies extremely fast even when the requested change looks small.
Worth watching closely before starting a larger development pass.