Build a Food Recommendation Agent with Doubao 1.6, Trae, and Volcengine MCP

Preface

Volcengine recently released Doubao large model 1.6, introducing a new AI‑cloud‑native paradigm: large model + Agent development platform.

The model can be accessed via Doubao 1.6, and the Agent platform can be used through PromptPilot for interactive prompt optimization and MCP protocol calls on Volcengine Ark, simplifying deployment.

I tried it by building a food‑recommendation assistant agent with Trae + Doubao 1.6 + Volcengine MCP.

1. New Features of Doubao 1.6

Doubao 1.6 includes three variants:

Doubao-Seed-1.6-thinking

Doubao-Seed-1.6

Doubao-Seed-1.6-flash

All three support multimodal input and a 256K context window.

The model now supports deep reasoning, multimodal understanding, and GUI operations, meaning it can both see and act.

Inference speed, accuracy, and stability are significantly improved, enabling more complex business scenarios.

In a benchmark for the 2025 national college entrance exam math paper, Doubao scored 144, ranking first nationwide.

Doubao 1.6 offers “edge‑thinking‑while‑searching” and DeepResearch capabilities, allowing autonomous research, planning, and tool usage.

DeepResearch can turn multi‑day, multi‑person research reports into 5‑30 minute summaries with automatic information extraction.

The entire series natively supports multimodal reasoning, powering real‑world problem solving.

This capability underpins real‑time video call features in the Doubao app and can be applied to e‑commerce review, autonomous driving annotation, safety inspection, and more.

2. Volcengine MCP

Volcengine MCP Hub integrates official cloud services and third‑party tools, supporting quick navigation to Ark or other MCP‑compatible platforms (Trae, Cursor, Python, etc.) with Remote and Local deployment options.

Through the Trae + Ark model service + MCP Market, Volcengine provides an end‑to‑end solution for large‑model application deployment.

The MCP Marketplace is available at https://www.volcengine.com/mcp-marketplace.

MCP services cover compute, storage, databases, containers, middleware, big data, security, search tools, and more, with both local and cloud deployment options.

3. How to Build a Food Recommendation Assistant

Locate the veFaaS MCP in the MCP Servers (https://www.volcengine.com/mcp-marketplace).

Open the veFaaS MCP detail page.

In the AI Idea tab, select Trae, choose the JSON(URL) option, and click “Generate”.

Click “Go to Trae Configuration”, which opens the domestic Trae UI and automatically loads the MCP tab.

Paste the generated JSON string and confirm to add the veFaaS MCP.

Create a new agent named “Food Recommendation Agent”.

Use the following prompt (shown in a code block) to define the agent’s behavior:

# Food Recommendation Agent Prompt

## Core Functions
1. Receive user‑specified city/region name
2. Recommend 10 classic dishes for that area
3. Generate a complete HTML file with all recommendations

## HTML Requirements
- Responsive design for all devices
- Card‑style layout for each dish
- Hover animations and basic styling
- Structure: navbar, recommendation area, footer

## Dish Information
- Name
- Description (2‑3 sentences)
- Budget price
- At least three tags (e.g., Cantonese, Traditional, Breakfast)
- Recommendation reason
- Generation timestamp displayed
- No dish images required

## Output Format
Output only the full HTML code without any extra explanation.

After saving, the agent is created.

The prompt was generated by Doubao‑Seed‑1.6.

In Trae’s dialog, select Doubao‑Seed‑1.6 and input:

I need to develop a food recommendation assistant agent, generate the prompt for me

Trae returns a professional prompt that can be further refined.

4. Demo

Switch back to Trae, select Doubao‑Seed‑1.6 and the “Food Recommendation Agent”.

Enter: Recommend 10 foods in Chengdu Trae generates a chengdu_food_recommendations.html file containing dish names, prices, categories, features, and recommendation reasons.

Open the HTML file in Chrome to view the results.

The assistant displays the recommended dishes with name, price, category, features, and reasons.

5. Volcengine PromptPilot

PromptPilot is a platform that helps developers build large‑model applications by constructing, debugging, and optimizing prompts through user feedback and data‑driven mechanisms.

It provides online bad‑case detection, continuous runtime optimization, and a closed‑loop workflow covering prompt engineering, AI engineering, fine‑tuning, solution output, and task‑specific evaluation.

Key capabilities include interactive intent extraction, automatic diverse answer generation, automated self‑reflection and optimization, and continuous prompt iteration as models and requirements evolve.

Example: optimizing a simple “generate a navigation website” prompt into a more professional version (see screenshot).

Conclusion

Doubao 1.6 offers stronger reasoning, multimodal understanding, and AI programming capabilities.

Pricing (illustrated below):

For most enterprise use (0‑32K tokens) the cost is ¥0.8 per million input tokens and ¥8 per million output tokens; 32K‑128K tokens cost ¥1.2 input and ¥16 output; 128K‑256K tokens cost ¥2.4 input and ¥24 output.

The large‑model + Agent platform embodies the AI‑cloud‑native paradigm, providing end‑to‑end support from inference optimization to secure deployment, enabling production‑grade agents at scale.