Unveiling Andrew Ng’s Four AI Agent Design Patterns with Coze

At the recent Sequoia AI event, Andrew Ng presented four primary AI agent design patterns—Reflection, Tool Use, Planning, and Multi‑agent Collaboration—and showed how they can be quickly realized on the Coze platform.

1. Reflection pattern

Meaning: The agent reviews and corrects its own output.

Background: Large language models sometimes generate incomplete responses; the Reflection pattern lets the model iteratively self‑optimize to improve quality.

Scenario: Generating an industry short commentary. The model writes a first draft, reads it, identifies needed edits, and rewrites, repeating as necessary.

Workflow:

Step 1: Start node receives user input.

Step 2: LLM node generates a short commentary using real data, case studies, and a SWOT model.

Step 3: The generated result is fed back as a new prompt for the LLM to review and refine.

Step 4: Output the final commentary.

First iteration produces a concise overview with real data; the second iteration adds detailed SWOT analysis and case examples, yielding a richer short report.

2. Tool Use pattern

Coze natively supports extensive tool calls, enabling the agent to generate code, invoke APIs, and perform other concrete actions.

3. Planning pattern

Meaning: The agent decomposes complex tasks and executes them according to a plan.

Background: LLMs can hallucinate or rely on outdated training data; Planning guides the model through a structured, multi‑step process to improve output quality.

Scenario: Producing a high‑quality research report. The agent first searches the web, filters results, extracts key information, and then synthesizes a concise report.

Workflow:

Step 1: Start node receives user input.

Step 2: Browser plugin searches the keyword and returns ten relevant website links.

Step 3: Headline search fetches detailed content from those links.

Step 4: LLM scores each page; scores above 3 are recommended for citation.

Step 5: Agent notifies the user that reading is complete and summarisation is in progress.

Step 6: LLM processes high‑scoring content, applying real data, case studies, and SWOT analysis.

Step 7: Output the final report.

Resulting example: a short commentary on the milk‑tea industry, showing the agent’s ability to browse, extract, and summarise information.

4. Multi‑agent Collaboration pattern

Goal: Multiple specialized agents cooperate to complete a task.

Three expert agents are defined for travel planning:

Destination recommendation expert – uses search tools to suggest locations based on user needs.

Flight & hotel expert – queries ticket and hotel services to propose suitable options.

Itinerary planning expert – integrates outputs from the other agents, creates a full itinerary, and exports it to PDF.

Workflow steps:

Step 1: Place the three agents on the canvas and set hand‑off conditions.

Step 2: Start the conversation, allowing agents to exchange information.

The resulting travel plan demonstrates high‑quality, multi‑agent output.

Comparison of the three modes