Labmate AI

Inspiration

Scientific and academic research often involves repetitive, time-consuming workflows: finding relevant information, going through documents, extracting evidence, organizing findings, and turning them into useful conclusions.

We wanted to explore a different approach to AI—one that does more than simply answer questions.

That led us to build LabmateAI, an autonomous AI research partner designed to help users execute complex scientific research workflows.

What LabmateAI Does

LabmateAI takes a research task from the user and turns it into an actionable workflow.

Instead of relying on a simple question-and-answer interaction, the agent can:

Understand and break down a complex research task
Plan the steps required to complete the task
Retrieve relevant information and sources
Use available tools to perform research actions
Analyze and organize the collected information
Generate a structured, evidence-backed result
Maintain relevant context throughout the workflow

The goal is to move from "AI that answers" toward "AI that helps execute the work."

How We Built It

LabmateAI was built around Google's agentic AI ecosystem.

The core architecture combines Gemini with a Google agent framework and Google Cloud services. The agent coordinates the research workflow, uses tools when required, retrieves relevant information, and produces the final structured output.

At a high level:

User → LabmateAI Agent → Planning → Tools/Retrieval → Analysis → Structured Result

The application is designed so that the agent, rather than the user, coordinates the individual steps of the workflow.

What We Learned

Building LabmateAI taught us that creating an agent is much more than connecting an LLM to a chat interface.

We learned about:

Agentic workflow design
Tool calling and multi-step task execution
Retrieval-augmented generation
Managing context and state
Connecting AI agents with cloud services
Designing reliable AI workflows
Evaluating AI-generated results
Building and deploying an end-to-end AI application

One of our biggest lessons was that giving an AI model more capabilities is not enough—the workflow around those capabilities needs to be carefully designed.

Challenges

One of the biggest challenges was making the system behave like an actual agent rather than a conventional chatbot.

We had to think about how the agent should decide which steps to take, when to use a tool, how to organize retrieved information, and how to present the final result clearly.

We also faced challenges around integrating the agent with external services, managing application state, and deploying the system reliably on Google Cloud.

What's Next

We plan to continue improving LabmateAI with better research workflows, stronger source verification, improved agent evaluation, additional tools, and more reliable long-running task execution.

Our long-term vision is to make LabmateAI a practical AI teammate for students, researchers, and scientific workflows.