┌─ // project ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐

roblox_bot

live
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
stackpython · claude api · yolo · opencv
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

Roblox Bot splits a real-time task across agents with very different latency requirements — a study in hierarchical agency.

The structure: three independent async tiers, each operating at its own cadence. Tier 1 is YOLOv8/OpenCV running at ~60fps — pure reaction, frame-level detection for movement and targeting. Tier 2 is Claude Haiku receiving state summaries every 1-2 seconds, producing tactical intent (engagement choices, positioning). Tier 3 is Claude Sonnet processing high-level game state on a longer loop, setting strategic objectives and adjusting macro-level goals.

The tiers don't wait for each other. Lower tiers proceed with the last available signal from higher tiers, which keeps cascading latency from becoming a blocking problem. The pattern is general — it applies to any real-time task that benefits from layered reasoning.

────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

// highlights

·

YOLOv8/OpenCV reactive layer at ~60fps

·

Claude Haiku for tactical decisions at ~1-2 second intervals

·

Claude Sonnet for strategic goals at ~5-10 second intervals

·

Fully async multi-process architecture — tiers don't block each other

·

Custom game state extraction from screen capture

·

Prompt engineering optimized for spatial and tactical reasoning