Product Feature Research
wiki / ux-patterns / Agent Reasoning Display

Agent Reasoning Display

The problem

Users want to see what the agent is doing while it builds — for trust, for interruption / stop points, and for learning. Too little = opaque "AI magic." Too much = overwhelming chain-of-thought dump.

Observed implementations

Anything.com — Named reasoning stages

From Anything.com - Agent Chat: - Each reasoning step has a named heading ("Considering Platform Implementation", "Prioritizing Web Implementation"). - Icon prefix: Anything logo + small sparkle. - Prose body (~2-4 sentences). - No tool calls shown during reasoning — the "thinking" is prose, not code execution.

Replit — Inline tool-call blocks

From Replit - Agent Chat with Preview: - Natural-language status ("Now let me build and sort both workflows"). - Each action is a collapsible tool-call card: - ▸ Ran cli (expandable — shows the command + output). - ▸ 4 actions (grouped — rolls up multiple sub-steps). - Agent takes screenshots to verify UI and includes them in the thread. - Narration quotes actual runtime values ("4 total invoices, 2 pending review, 1 approved...").

Comparison

Dimension Anything.com Replit
Visibility granularity Stage-level (high-level) Tool-call-level (granular)
Content Prose reasoning Tool calls + outputs
Shows actual data No (just reasoning) Yes (quotes real values from run)
Expandability Static text Collapsible cards
Screenshots No Yes (agent captures UI)
💭

Anything.com emphasizes the why (named reasoning stages, no tool mechanics). Replit emphasizes the what (tool calls, real values, screenshots). These map to audience:

  • Anything's indie-maker ICP doesn't need to see pnpm install; they need to know the agent considered their options.
  • Replit's dev-adjacent ICP actually benefits from seeing the commands — developers evaluate agents on the rigor of the tool loop.

Both are valid. Neither would serve the other's audience as well.

Hybrid potential

A future iteration could blend both: - Reasoning stages (for the why). - Expand-on-demand tool calls (for the what). Most agents (Cursor, Claude Code) are converging in this direction.

Sources