Designing for Agents (為 Agent 設計產品 - 英文原文版)

原始來源與檔名：20260512_2026-04-28T092735+0800-Designing for Agents.md

來源：[[@teddy_riker]] / X (Twitter) — 2026-04-23 原始檔名：2026-04-28T092735+0800-Designing for Agents.md

NAPKIN | 餐巾紙

餐巾紙公式

(This is the original English text of Article #129. The core concepts are identical.) Old Paradigm: User -> Interface -> Database New Paradigm: User -> User’s Agent -> Software’s Agent -> Database Design Pillars: 1. Teach Agents (MCP Specs) + 2. Feedback Loops (Rationale parameters) + 3. Mind Context Gaps (Exchange context, not exact IDs).

This article by Teddy Riker outlines the paradigm shift from human-first UI to agent-first APIs (Headless software). As agents like Claude and ChatGPT become the primary operators of software (as seen in Salesforce’s Headless 360), product teams must design for intermediaries. This means explicitly providing markdown specs via MCP, requiring ‘rationale’ logs to build feedback loops, and designing APIs that ask for context rather than exact system codes.

一句話

本文是第 129 篇文章《为 Agent 设计产品》的英文原始出處。作者 Teddy Riker 深刻闡述了在 AI Agent 接管軟體操作的未來，產品介面將從「人機互動 (GUI)」轉向「Agent 對 Agent」的無頭 (Headless) 互動。開發者必須主動將操作規範交給 Agent（如 Notion 的做法）、建立基於調用理由 (Rationale) 的反饋迴圈，並優化 API 以處理上下文缺口（Context Gap），讓軟體真正具備 Agent 親和力。

餐巾紙草圖

[ Agent-to-Agent Architecture ]

Diego (Human)
  |
  v
Chief of Staff Agent (Has Context: Calendars, Slack, Receipts)
  |
  |--- "Here is the context of the expense: Team dinner at Kokkari." --->
  |
Expense System Agent (Has Rules: GL Codes, Company Policies)
  |--- Translates context to GL Code 6400 --->
  v
Database Update.

ROUND 1: SKELETON | 骨架掃描

“這篇文章在說什麼”

• The Shift: The UI is not dead, but the 80/20 rule has flipped. 80% of software interactions will be executed by agents. Salesforce’s Headless 360 is proof that the old UI moat is eroding.
• The New Pattern: User -> User's Agent -> Software's Agent -> Database.
• Three Rules for Agent Design:
  1. Teach Agents to Succeed: Provide explicit documentation (like Notion’s Markdown spec) directly to the agent via MCP when needed. Don’t make them guess.
  2. Build Feedback Loops: Require a rationale parameter for all tool calls to understand intent, and create specific tools for agents to report blockages. Agents are more consistent in feedback than humans.
  3. Mind the Context Gap: Understand that the User’s Agent holds personal context (emails, calendars), while the Software’s Agent holds business rules. Design interactions that exchange context to bridge the gap, rather than forcing the user to map data manually.

ROUND 2: DISSECTION | 血肉解剖

“憑什麼這麼說”

核心論證鏈

1. Erosion of UX as a Moat: Familiarity with complex enterprise UIs used to lock in customers. Once an agent abstracts the UI away, functionality and API/MCP compliance become the only true competitive advantages.
2. Intent-driven API: Traditional APIs demand specific IDs or enums (like a GL Code). Agent-driven APIs thrive on natural language context. It is fundamentally a shift from deterministic function calls to probabilistic context resolution.

關鍵證據

• Notion vs. Slack MCP implementations: Notion forces the agent to read its markdown rules resulting in perfect formatting, whereas Slack assumes default markdown knowledge, resulting in broken messages.
• Real-world observability at Ramp: Using rationale logs to discover new feature requests directly from agent failures.

邊界條件

• The success of this architecture heavily relies on the reasoning capabilities and adherence of the foundational LLMs. If an agent hallucinates a rationale or misinterprets context, the feedback loop could poison the product roadmap.

ROUND 3: SOUL | 靈魂提取

“還能怎么用”

• 知識連結: Directly aligns with the “Harness Engineering” concept discussed in earlier articles, where managing how the agent interacts with external systems (via MCP) is as important as the model itself.
• 深層洞見: “The interface used to sit between Diego and his expense system. Now it sits between his agent and yours.” This completely redefines the job of a Product Designer. You are now designing protocols for alien intelligences, not just intuitive buttons for humans.
• 行動呼籲: Review your product’s API or MCP tools. Ask yourself: “If an agent with no prior knowledge of our system calls this, what context does it lack?” Then, write that context directly into the tool’s description prompt.