How Dynamo AI’s Deep Research Agent Works

A Complete Guide For researchers, students, and power users who want to understand what’s actually happening under the hood.

What Is the Deep Research Agent?

Most AI chatbots answer in one shot. You ask a question, the model pulls from its training data, generates a response in a few seconds, and moves on. That works well for quick explanations, summaries, and casual questions.

The Deep Research Agent is a fundamentally different beast. Instead of a single response, it runs a multi-step autonomous pipeline — planning, searching, extracting, finding gaps, synthesising, and then cross-checking against real academic literature — before it writes a single word of the final report. The whole process takes 4–8 minutes, and that time is entirely intentional.

It’s designed for questions that deserve a serious answer: industry analysis, scientific topics, policy research, competitive landscapes, and literature reviews.

How It’s Different From the Other Modes

Research Mode chains Claude, Gemini, and GPT across one search to give you a strong single-pass answer. Deep Research Agent is a separate system entirely — it doesn’t use that pipeline at all. It runs its own multi-step loop.

The Pipeline, Step by Step

Here’s exactly what happens from the moment you hit Send in Deep Research mode.

Step 0: Authentication Gate

Before anything runs, the system checks that you’re on the Pro plan. Deep Research is computationally expensive — 5 AI inference calls, 6 web searches, and a Semantic Scholar API query all happen for one request. This isn’t an artificial restriction; it’s a resource reality.

Step 1: Research Planning

The agent doesn’t just search for your raw question. It first asks Gemini to design a research strategy.

It gives the model your query and asks it to generate exactly 6 distinct search queries covering these angles:

1. Overview and background
2. Current state and data
3. Key actors and stakeholders
4. Challenges and criticisms
5. Recent developments (2024–2026)
6. Future outlook and expert predictions

The model returns a JSON array of 6 strings. If the parsing fails (or the model returns fewer than 3 queries), the system falls back to a safe set of templated queries built from your original topic. This means the planning step never becomes a single point of failure.

Step 2: Six Parallel Web Searches

Each of the 6 planned queries is sent to the search engine one by one (with a short stagger to avoid rate limits). These aren’t generic web searches — each is targeted at a specific research angle, so the results cover genuinely different parts of the topic rather than returning 6 versions of the same Wikipedia summary.

The raw context from all 6 searches is collected and labelled (Overview, Current data, Key players, etc.) before being passed forward.

Step 3: Insight Extraction

The agent takes all 6 search results — potentially thousands of words of raw web content — and asks Gemini to act as a research analyst. The task: read everything and extract the 12 most important factual insights, statistics, and findings, numbered [1] through [12], each with a reference back to which source it came from.

This step is what separates the agent from a simple “search and dump” pipeline. The model is forced to read critically, not just collect.

Step 4: Gap Analysis

This is where the agent does something most research tools skip entirely. After extracting insights, it asks a separate Gemini call to act as a senior research critic and identify:

• 4–6 significant gaps in the existing literature
• Underexplored angles
• Open questions that haven’t been adequately addressed

These gaps are passed into the final synthesis prompt, which means the final report doesn’t just summarise what’s known — it explicitly tells you what isn’t known or hasn’t been studied enough.

Step 5: Full Report Synthesis

Now the agent writes the actual report. It sends Gemini a structured prompt containing:

• The 12 numbered insights with citation markers
• The 4–6 research gaps
• Up to 6,000 words of the original raw web context (as a factual backstop)
• A strict 11-section structure requirement

The required sections are:

1. Executive Summary (150 words)
2. Introduction & Background
3. Current State of the Field
4. Key Findings (with inline [1], [2] citations)
5. Key Stakeholders & Actors
6. Challenges & Criticisms
7. Recent Developments (2024–2026)
8. Research Gaps & Open Questions
9. Future Outlook & Expert Predictions
10. Conclusion
11. References

The model is instructed to write a minimum of 1,500 words and to ensure every claim is supported by one of the numbered citations from Step 3. This grounds the report in the actual web research rather than the model’s training data.

Step 6: Semantic Scholar Sources Matrix

Once the report is written, the system makes a call to the Semantic Scholar Graph API — a free, openly accessible academic database with over 200 million papers. It searches for papers relevant to your original query and retrieves up to 8 results, each with:

• Full title (linked to DOI)
• Authors
• Publication year
• Citation count
• Abstract excerpt

These are formatted into a Sources Matrix table appended to the bottom of the report. This serves two purposes: it gives you real academic citations to follow up on, and it lets you sanity-check the report’s claims against peer-reviewed literature from the Semantic Scholar, which is a public endpoint, and the agent accesses it directly without any authentication.

The Live Activity Feed

While the agent runs, you see a live activity log inside the research card — timestamped messages like:

The frontend polls the backend every 3 seconds for status updates. The backend maintains a deduplicated activity log per job (in-memory, keyed by a short UUID), and the frontend only renders messages it hasn’t seen before — so the log grows cleanly without duplicates even across dozens of poll cycles.

The timer in the top-right of the card ticks up in real time, independent of the polling. It’s intentionally visible — deep research is supposed to take time, and the timer makes that feel purposeful rather than broken.

What You Can Do With the Report

Once the report is complete, the action bar at the top gives you six tools:

Edit — Opens the full markdown report in an inline text editor. You can rewrite sections, add your own notes, remove irrelevant parts, and fix formatting. When you save, the display re-renders from your edited version. Everything downstream (download, copy) uses your edited version.

Download .md — Saves the full report as a Markdown file with a filename derived from your research query. Useful for dropping straight into Obsidian, Notion, or any document editor.

Copy — Copies the full report to the clipboard as plain markdown text.

Save to library — Saves the report to your Document Library. Once saved, Dynamo AI will automatically inject a summary of the report into the context of future chats, so the AI knows you’ve already researched this topic.

Ask a follow-up → — Opens an inline follow-up pane directly inside the research card (no new chat needed). You can ask things like “add APA7 references”, “expand the challenges section”, “translate this to a formal academic tone”, or “write an abstract under 200 words”. The follow-up is sent with the full report as context and renders its answer in a new assistant bubble below.

Draft Academic Paper — Takes the report content and asks Gemini to rewrite it in a formal academic paper format with proper section headings, passive voice, citation notation, and a structured abstract.

Verify with Papers — Sends the report excerpt back through the Semantic Scholar pipeline and asks Gemini to cross-reference the 5 key claims against the retrieved papers, producing a structured evidence table showing which claims are supported, which are partially supported, and which need more evidence.

The Trust Layer — Verify on Any Message
Starting in the latest version, every substantive assistant message in the regular chat — not just Deep Research reports — shows a small Verify shield button. Clicking it runs the same Semantic Scholar cross-check, but for that specific message and question. The evidence panel opens inline within the message bubble. This means you can fact-check a Fast Mode answer just as easily as a Deep Research report.

How to Use It Effectively

Be specific in your query. “AI in healthcare” will produce a broad overview. “AI-assisted diagnostic imaging in radiology: current accuracy benchmarks and regulatory challenges” will produce a report you can actually use. The research planner generates better search angles when it has a precise topic to work with.

Let it finish. The 4–8 minute window is real. The agent is doing the equivalent of a human researcher’s half-day of reading and note-taking. Switching away and coming back is perfectly fine — the card stays in your chat session, and the poll will catch the result whenever you return.

Use the follow-up pane for refinement. The best workflow is: get the initial report → edit out any obviously irrelevant sections → ask a follow-up for the specific extension you need. This is faster and more accurate than trying to describe the perfect report upfront.

Save important reports to your library. Once saved, the report becomes a persistent memory. Future chats will have context about your existing research, so Dynamo AI won’t re-explain things you already know and will build on what you’ve already found.

Check the Sources Matrix. The papers Semantic Scholar returns are real, peer-reviewed work. If a key claim in the report appears in the Sources Matrix with a high citation count, that’s a strong signal that the claim is well-established. If the paper list looks unrelated to the report’s claims, that’s a signal to dig deeper or verify independently.

What It’s Not
Deep Research Agent is not a replacement for genuine academic research. It’s a first-pass intelligence tool. The web search results it reads may include non-peer-reviewed content. The AI synthesis may oversimplify complex debates. The Sources Matrix retrieves papers by keyword relevance, not by reading the papers themselves.

Use the reports as a structured starting point — a way to rapidly understand a field, identify the key debates, and find the real papers and sources worth reading in full. The Verify button and the Sources Matrix are there precisely to help you do that second step.

The Deep Research Agent is a Pro-tier feature on Dynamo AI. Fast Mode and Research Mode are available on Free and Plus, respectively.