Beyond the sparkle: giving thinking, tool calls, and retrieval their own icons
Glittering_Device653 · reddit · 2026-08-27
A designer sparked a discussion on the missing visual language of agent UIs: the sparkle has become the universal symbol for "AI" — on rewrite buttons, autocomplete, summarizers, and agents that spent 40 seconds calling six tools, all with the same sparkle, so users can't tell "thinking" from "tool call" from "retrieval".
His thesis: sparkle should mean only "a model generated this" — a provenance marker, not a feature badge. Retrieval, tool calls, handoffs, and guardrails each deserve their own shape so users can tell them apart when things go wrong.
Practical details:
- "Agent" is the hardest icon: a robot reads as a toy, a person reads as a user (exactly wrong); he settled on a consistent ring
- He renders every icon to a 24px bitmap and computes pairwise overlap — 19 pairs were essentially indistinguishable at real usage size (compile/terminal/shell/run script all converge to the same shape)
- Some concepts may be undrawable: chunk overlap, context budget, attention head, best-of-n, ablation
He open-sourced his icon set IconMind (MIT, free, no paid tier).
More from coding & agent
- Grok Build Agent: Autonomously Fixes Errors via Loops — tetsuoai · 2026-08-27
- Google Killed Gemini CLI Free Tier: Available Alternatives — Agreeable_Craft_8943 · 2026-08-27
- User finds Codex output readability superior to Claude Code — auto_grad_ · 2026-08-27
- Designing MemoryService for Multi-Agent Systems: Policy or Storage Layer? — Fun-Following-1723 · 2026-08-27
- Rabbit R1 turned into a desktop companion with auto-chat features — SimonBalmain · 2026-08-27
- Open source tool 'cat': VSCode-style agent multiplexer with auto-save — jasonkneen · 2026-08-27