Engineering
The Engineering Story: Managing Agents with Guardrails
Cati is a prime case study in governing autonomous AI development with strong verification gates. Over a six-week period, the repository recorded 255 commits, 27 PR merges, and 65 Go test files across 38,200 lines of Go.
1. The Geometry Snapping Trap
In late September, user options for width, height, aspect behavior, and padding had drifted into conflicting rules between the static image pipeline and the media player. Agents working on aspect ratio fixes inadvertently caused pixel-aspect distortions on non-square fonts.
The Fix: Rather than patching ad-hoc formulas, we specified a centralized Geometry Planner (Issue 075). The team created an interactive aspect test matrix with 24 test cases. The planner snaps to native subcell lattices, enforces alignment padding, and scales row heights relative to terminal character metrics.
2. High-Performance Closed-Form Sextant Math
Rendering high-resolution sextant glyphs originally used brute-force per-mask pixel
scoring, which choked frame rates on terminal video playback. Working with
google-labs-jules[bot] and local Harnez workers, the scoring loop was replaced with a
closed-form error calculation and lookup tables, coupled with early exits for
sparkline candidate searches.
Measured Result: A 40% reduction in sextant render latency and a 27.3% latency reduction for S2 mode. Strict parity benchmarks guaranteed zero visual regressions.












