Inspiration
Researching a complex topic can be frustrating and time-consuming. We often have to search through dozens of websites, papers, reports, and articles, then compare different opinions and decide which information is actually reliable.
We wanted to build something beyond a traditional AI chatbot. Instead of simply asking AI a question and receiving an answer, we wanted an AI system that could plan the research, investigate the topic, check its own findings, and explain the evidence behind its conclusions.
That idea became DeepResearch Agent — Autonomous Research System.
What it does
DeepResearch Agent turns a complex research question into an autonomous research workflow.
A user can ask something like:
"Will AI replace software developers by 2030?"
The system then:
Creates a research plan Breaks the question into smaller topics Finds relevant sources Analyzes evidence Extracts important claims Identifies conflicting information Critically reviews the findings Performs additional research when needed Generates an evidence-backed report Stores the research for future use
The system uses multiple specialized agents that work together:
User ↓ Research Manager ↓ Search Agent ↓ Evidence Agent ↓ Critic Agent ↓ Additional Research ↓ Report Agent ↓ Final Research Report
The goal is to make research feel less like asking a chatbot and more like working with an AI research team.
How we built it
We built DeepResearch Agent using Gemini and Google ADK to create and coordinate the specialized agents.
Our architecture uses:
Gemini — reasoning, analysis, and report generation Google ADK — multi-agent orchestration Firestore — storing research, sources, claims, and agent state Cloud Run — deploying the application Pub/Sub — supporting asynchronous/background research workflows React — the user-facing research dashboard
The Research Manager acts as the coordinator. It decides what needs to be researched and when additional investigation is necessary.
The Critic Agent is especially important because it doesn't simply accept the generated findings. It looks for weak evidence, contradictions, unsupported claims, and gaps in the research.
Challenges we ran into
One of our biggest challenges was making the agents work together as a real system rather than simply having several independent AI prompts.
We had to carefully define:
What each agent is responsible for What information should be passed between agents When more research is necessary How to handle conflicting evidence How to store research state How to prevent the system from becoming overly confident
Another challenge was balancing autonomy and reliability. We wanted the system to make decisions on its own, but we also wanted it to recognize when the available evidence was insufficient.
Accomplishments that we're proud of
We're proud that DeepResearch Agent goes beyond a simple AI question-and-answer experience.
Our system can:
🧠 Plan its own research 🤖 Coordinate multiple specialized agents 🔎 Discover and analyze sources 📚 Connect claims with evidence 🧐 Challenge its own findings ⚠️ Detect conflicting information 🔄 Perform additional research when necessary 💾 Remember previous research 📊 Produce structured research reports
Most importantly, we built the project around the idea that AI should not only generate information — it should also question and verify it.
What we learned
Building DeepResearch Agent taught us that creating an agentic application is very different from creating a normal chatbot.
A basic chatbot follows:
Question → Answer
Our system follows:
Question ↓ Plan ↓ Research ↓ Analyze ↓ Challenge ↓ Verify ↓ Improve ↓ Report
We learned that giving every agent a clear responsibility makes the overall system more reliable and easier to develop.
We also learned that evidence and uncertainty matter. A confident AI answer isn't necessarily a correct answer, so the ability to challenge conclusions is an important part of our system.
What's next for DeepResearch Agent — Autonomous Research System
We want DeepResearch Agent to become a long-term research partner, rather than something users only open once.
Our next steps include:
🔔 Continuous monitoring of research topics 🆕 Automatic detection of important new sources 📈 Tracking how conclusions change over time 🧩 More advanced evidence and relationship graphs 👥 Collaborative research between users and AI agents 🌐 Support for more research domains and languages 📱 A more polished and accessible research dashboard
Our long-term vision is simple:
DeepResearch Agent should not just tell you what it thinks. It should investigate, question, verify, remember, and show you why it reached its conclusion.
