Designing for Agents (為 Agent 設計產品 英文原版)

原始來源與檔名：Designing-for-Agents-英文版.md

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

(註：此文為「為 Agent 設計產品」的英文原文版，核心知識點已提取於中文版筆記中。)

NAPKIN | 餐巾纸

餐巾紙公式

新世代產品架構 = User -> User’s Agent -> Software’s Agent -> Database

The 80/20 rule of UI is flipping: soon 80% of software interaction will be through AI agents via MCPs. Software must expose capabilities with explicit markdown rules (like Notion does), require “rationales” for tool calls to build feedback loops, and bridge the context gap between a user’s personal agent and the enterprise software’s agent.

一句話

The interface used to sit between the user and your system. Now it sits between their Agent and your Agent.

餐巾紙草圖

[Old World]
User ---> (UI Clicks) ---> Database

[New World: Agent to Agent]
User ---> (Voice/Chat) ---> User's Agent (Claude) 
                                | (MCP / API) 
                           Software's Agent (Notion/Salesforce)
                                |
                             Database

Design Rules:
1. Provide MD Specs explicitly (Don't let agents hallucinate formatting).
2. Require 'rationale' in tool calls for intent observability.
3. Bridge Context Gaps (User Agent has Calendar, System Agent has Policies).

ROUND 1: SKELETON | 骨架掃描

“這本書在說什麼”

• 核心問題: With the rise of AI agents, traditional GUI is becoming a secondary interaction layer. How should product teams design software for non-human, Agent users?
• 核心答案: Embrace “Headless” architecture via MCPs. Provide explicit documentation (Markdown specs) directly to the calling agents, enforce rationale logging to reconstruct user intent, and elegantly resolve the “Context Gap” between disparate AI systems.
• 論證結構: Acknowledges the trend (Salesforce Headless 360) -> Defines the new interaction pattern -> Teaches how to help agents succeed (Notion vs Slack MCP) -> Highlights the need for feedback loops (Rationale parameter) -> Illustrates the context gap (Expense reporting example).

章節骨架

1. The Trend: Salesforce’s Headless 360 proves UI is losing its moat. Ramp sees a 10x surge in MCP traffic.
2. New Interaction Pattern: Transition from User -> UI -> DB to User -> User's Agent -> Software's Agent -> DB.
3. Teach Agents to Succeed: Notion explicitly forces agents to read its markdown spec; Slack leaves agents guessing, leading to broken formatting.
4. Build Feedback Loops: Agents using APIs don’t pass along chat context. Forcing a rationale parameter allows product teams to reverse-engineer user intent and build better macro-tools (like incident report generators).
5. Mind the Context Gap: When an individual’s agent (which knows their calendar/email) talks to an enterprise agent (which knows GL codes and policy), they must combine context rather than dumping the decision back on the human.

ROUND 2: DISSECTION | 血肉解剖

“憑什麼這麼說”

• 隱形假設: The author assumes that MCP (Model Context Protocol) will become the ubiquitous standard and that LLMs will reliably follow the injected specs and rules provided by the software agents.
• 邊界條件: This “headless” approach is highly effective for data-driven, workflow-heavy SaaS (CRMs, Notion, Expense tracking). It is not applicable to spatial, highly visual, or real-time creative tools (e.g., Video editing, 3D modeling) where GUI is intrinsic to the workflow.

ROUND 3: SOUL | 靈魂提取

“還能怎么用”

• 知識連結: Headless Architecture, API-First Design, Observability, Model Context Protocol (MCP).
• 深層洞見: Designing for agents removes the “human error” of UI navigation but introduces “LLM hallucination”. The ultimate UX for an AI agent is deterministic, well-documented boundary conditions injected Just-In-Time.
• 行動呼籲: If you expose an API or MCP tool, immediately add a required rationale: string parameter to your schemas. It will give you unprecedented visibility into what your users actually want to achieve.

Architectural Deep Dive: Headless Agent Interaction

前言/背景

The Model Context Protocol (MCP) has initiated a paradigm shift. Users are increasingly delegating tasks to personal AI agents, meaning enterprise software must pivot from designing graphical interfaces for human eyes to designing robust, intention-aware APIs for LLMs.

章節詳細總結

1. The Headless Paradigm Shift

The launch of Salesforce Headless 360 validates that the traditional monolithic UI is no longer a defensible moat.

• Architecture: The communication chain has evolved. The new standard is User -> User's Agent -> Software's Agent -> Database. Products must function as capable “Software Agents” that securely and efficiently process requests from external LLMs without relying on human point-and-click navigation.

2. Contextual Guardrails (The Notion Model)

Expecting an LLM to guess system-specific syntax is an anti-pattern that leads to hallucinations.

• Architecture: Notion’s MCP employs “Just-In-Time Context Injection”. It forces the calling agent to fetch an enhanced-markdown-spec before executing write commands. This programmatic constraint eliminates formatting errors and guarantees that the input matches the system’s database schema requirements.

3. Observability and Intent Reconstruction

In a headless world, traditional UX tracking (clickstreams, heatmaps) is blind.

• Architecture: Ramp solved this by enforcing a rationale parameter on every tool call. By logging and analyzing the LLM’s explanation for why it called a specific tool, the product team can reconstruct the user’s overarching intent, bridging the observability gap and surfacing ideas for higher-order macro tools.

4. Bridging the Context Gap

Agent-to-agent communication often suffers from asymmetric information.

• Architecture: A well-designed system does not push choices (e.g., selecting a GL code) back to the user. Instead, the User’s Agent (providing personal context like calendars) and the Software’s Agent (providing corporate policies) negotiate to deduce the correct outcome automatically, achieving zero-friction data processing.

總結與結論

• Philosophy: The best “UI” for an AI agent is explicit rules, injected exactly when needed.
• Future Outlook: Product teams that obsess over the developer/agent experience (DX) of their MCP servers—by providing clear constraints and capturing intent via rationale logging—will win the headless era.