Agent Ergonomics

How well 7 tools serve an agent

One reference agent, the same task per category, five trials each. Graded against deterministic ground truth — did the output actually render, and does it hold up when repeated? Where a tool was taught more than one way (official docs vs a skill), the sources are compared head-to-head.

7 of 7 tools · headline = the tool's best measured teaching source

hostile friendly
plantuml via skillA−diagrams+0.35+0.62+0.58+0.38+0.48+0.490+0.50
mermaid via official docsA−diagrams+0.23+0.62+0.63+0.36+0.35+0.480·
d2 via official docsA−diagrams+0.26+0.69+0.45+0.42+0.37+0.470·
matplotlib via skillA−data-visualization+0.39+0.59+0.33+0.25+0.15+0.380+0.15
react-email via skillB−advertising-creative+0.34+0.45+0.35+0.41+0.20+0.391·
excalidraw via official MCPB−diagrams-0.18+0.25+0.19-0.12+0.05+0.055·
latex via skillFdocuments+0.32+0.14+0.05+0.35+0.17+0.203·

Grades are anchored in a deterministic checker where one exists (renders? requirements met? reliable when repeated?) — a green matrix cannot rescue output that does not work. Surface columns are −1…+1 instrument readings.

Quality × cost

up-left is better · ◆ = on the frontier

Each point is one subject: its grade quality against the tokens an agent spends per successful run. The axes stay separate — a cheap unreliable tool and an expensive reliable one are different answers, not one number.

4k13k40ktokens / successful run (log)00.51qualityplantumlmermaidd2matplotlibreact-emaillatex