Deep Research

⬇️ Preview

Introduction

I recently built a deep-research agent on the Tencent Cloud Agent Development Platform (ADP). The project reshaped how I think about low-code workflow orchestration, so I’m sharing the process in case it helps anyone exploring AI agent development.

Project Background

Deep research goes far beyond keyword lookup. It breaks a topic into multiple dimensions, gathers information step by step, and composes a structured research report. That level of complexity is a perfect match for a workflow-driven approach.

Architecture Overview

Workflow hierarchy

The solution relies on nested workflows, each with a clear responsibility:

Core Flow Implementation

1. Topic breakdown and subtopic generation

After the user submits a topic, the workflow proceeds as follows:

Input topic -> Topic confirmation -> Generate four subtopics -> Subtopic confirmation

The key is to produce exploratory subtopics instead of blunt keyword permutations. For example, a study on "Artificial Intelligence" might yield:

• Current state and trends
• Real-world application cases
• Ethics and societal impact
• Future challenges and opportunities

Topic collection

Subtopic collection

2. Parallel batch execution

The batch node runs subtopic research in parallel with a concurrency of three—a balance between throughput and API rate limits.

3. Intelligent tool routing

When a step executes, the workflow selects the most suitable tool:

Dynamic routing ensures each information need leverages the right channel.

Key Technical Highlights

1. Model selection strategy

Choosing specialized models for each task keeps quality high while controlling cost:

2. Variable aggregation tricks

Variable aggregation makes it easy to merge results from multiple branches:

// Aggregation sample
{
  "result": [searchResults, paperResults],
  "keyword": [searchKeywords, arxivKeywords]
}

3. UX refinements

Progress messaging

Key stages display reassuring status updates:

• "Generating research steps for {topic}..."
• "Compiling the final report—this takes about one minute..."

Confirmation gates

Long-running actions include a confirmation step that:

• Shows the estimated execution time
• Lets users cancel before resources are consumed

Output Innovations

PDF delivery workflow

Beyond plain text, the workflow exports polished PDFs:

EdgeOne’s static hosting powers instant sharing of the finished report.

Summary and Next Steps

Potential areas to explore next:

1. Knowledge graph enrichment to map relationships between research themes.
2. Multimodal inputs by incorporating images, video, or audio analysis.
3. Collaborative research so teams can co-create reports in real time.
4. Incremental updates that let users append fresh findings to existing work.

I hope this walkthrough sparks ideas for your own AI agent experiments. Low code does not mean low capability—the magic lies in thoughtful orchestration. If you run into challenges, drop a comment and let’s compare notes.