The Agent Deep Dive: David Zhang’s Open Deep Research
This blog post explores David Zhang’s open-source framework for iterative AI research. We look at the system’s query process, parallel search executions, and markdown report creation, and compare it to both OpenAI’s and Perplexity’s research solutions.
AI-driven research is evolving fast. While Google pioneered Deep Research, OpenAI popularized it and Perplexity and Grok followed, open-source alternatives are just as exciting and help to understand how one of the first forms of autonomous useful agents works. In this post, I take a closer look at the inner workings of open deep-research by David Zhang (dzhng)—a lightweight, transparent framework for iterative AI research.
At under 500 lines of code, deep-research packs a surprising punch. It runs recursive research loops, refining queries based on the information it finds—no hidden logic, no unnecessary dependencies, just a clean, modular setup that’s easy to hack or extend.
I decided to put it to the test. With Langfuse, I traced its execution (thanks to @hassiebp for instrumenting it in this fork), broke down its decision-making, and measured performance. Below, I’ll show how it works, why it’s worth a look, and how it compares to other AI deep research tools.
Why I Like This Agent
- Simple – Minimalist design, easy to modify, and fully open-source.
- Iterative Research – Runs deep research loops, refining its queries at each step.
- Control Depth & Breadth – Users define how far (depth) and wide (breadth) the research should go.
- Parallel Execution – Fires off multiple queries at once, speeding up research.
Tech Stack:
- o3-mini as the language model for SERP generation, learnings, and the final report.
- Firecrawl for executing SERP queries.
Despite its minimal footprint, the framework is robust enough to generate meaningful follow-up questions, refine its search, and compile a structured final report. This simplicity makes it an approachable playground for anyone looking to experiment with or build upon an iterative AI research agent.
How Does the Agent Work?
To see how deep-research holds up, I asked it a more nuanced question:
“I don’t get how Whisper is fully open source (MIT), yet ChatGPT’s speech-to-text is vastly better than Perplexity’s and Apple’s. Can you help me figure this out?”
This isn’t a simple fact lookup—it requires synthesizing multiple sources, filtering insights, and producing structured responses.
With a breadth of 4 and a depth of 2 (which are the two configuration options that are available), we might see the agent generate multiple parallel queries—and then refine those queries after it sees what the search turned up. You’d end up with a thorough markdown report summarizing everything, plus any missing links the agent found intriguing.
Each iteration of this loop can be viewed as an event within Langfuse. Tracking these events provides a step-by-step replay of the agent’s logic—useful for debugging and fine-tuning. Here’s the public trace in Langfuse:
Note: The chart above is a Gantt-like overview of query processes.
Next, let’s break down each step of the research loop in more detail.
1. Initial Query
The system message and user prompt are created. This is where the agent sets the context.
2. Follow-Up Questions
The agent proposes up to three follow-up questions based on the initial prompt.
3. Search Queries
Based on those follow-up questions, the agent generates SERP queries to investigate the topic further.
4. Search Engine Calls
The system then spawns multiple queries to a search API (in this case, Firecrawl).
5. Learnings from the Content
After retrieving the search results, the agent processes them and generates up to three “learnings” (insights). It also proposes new follow-up questions, continuing the loop if necessary.
6. Markdown Report
Finally, once the maximum research breadth (e.g., 4) and depth (e.g., 2) are reached, the agent compiles a Markdown report summarizing key findings. It uses all the intermediate “learnings” from previous steps to form a coherent answer.
Thanks to the instrumentation we added above, all these steps are tracked in Langfuse—timestamps, search calls, LLM calls, and final outputs—providing a transparent view of how the agent arrives at its answer.
Comparison With Other AI Research Tools
Although deep-research is a small, open-source project, I wanted to see how it fares alongside two well-known AI tools: OpenAI’s ChatGPT (with advanced research features or higher usage tiers) and Perplexity Pro.
To keep it fair, I ran the same question on all three:
David Zhang’s Open Deep Research
- Time to finish: ~3 minutes and 34 seconds
- Estimated model usage cost: ~$0.71
- Generated report length: ~12k characters
- Trace: Langfuse Link
OpenAI’s Deep Research
- Time to finish: ~6 minutes
- Insights: ~27 sources
- Generated report length: ~87k characters
- Costs: Hard to pin down exactly; it likely depends on o3 and crawler usage.
- Example link: OpenAI Example Thread
Perplexity Pro Deep Research
- Time to finish: ~3 minutes
- Insights: ~49 sources
- Report length: ~8k characters
- Link to Perplexity Thread: Perplexity Deep Research Example
Takeaways
deep-researchis impressively agile for a lean, open-source tool.- Perplexity’s product can pull in a larger number of sources quickly, leveraging its specialization on web search.
- OpenAI’s Deep Research produced a lengthy report, utilized more sources, and conducted a thorough investigation based on follow-up searches. Overall, the reasoning capabilities (o3, post-training for this use case) are significantly stronger than those of the other contestants.
Ultimately, if you want transparency, low cost, and hackability, David Zhang’s deep-research stands out. If you need huge reports or integrated browsing at scale, you may prefer a more mature commercial solution.
Other Open Source Research Agent Frameworks
If you’re interested in exploring similar projects, check out these options:
What’s Next?
I will do more deep dives on interesting OSS agents to dissect how they work and what’s interesting about them. Email me if you have suggestions: [email protected]