Skip to content
~/rahultank
all projects

Scout — Autonomous AI Research Agent

An autonomous AI research agent that decomposes a question, searches the live web, and streams a synthesized, source-aware answer in real time.

  • TypeScript
  • React 19
  • LangGraph
  • Gemini
  • Tavily
  • Redis
  • SSE

A research agent that plans before it answers: it decomposes a question into sub-questions, runs live web searches, reads the results, and streams a synthesized answer in real time. LangGraph orchestrates the plan → search → synthesize loop, Gemini 2.5 Flash does the reasoning, and every step is pushed to the UI over SSE so the agent's thinking is visible as it works.

// Problem

A single LLM call answers from frozen training data — it can't research a current question, follow leads, or show its work. I wanted an agent that plans, searches the live web, and stays transparent about how it reaches an answer.

// Approach

Model the task as a graph instead of one prompt: decompose the question into sub-questions, run a web search per branch, read and rank the sources, then synthesize a single answer — streaming each stage so the reasoning is visible rather than hidden behind a spinner.

// Architecture

LangGraph orchestrates the plan → search → synthesize loop with Gemini 2.5 Flash as the reasoning model and Tavily for live web search; Playwright handles page extraction and Redis backs run state and caching. The React 19 frontend consumes a single SSE stream, rendering intermediate steps and the final answer token by token. Deployed on Vercel.

// Outcome

A public, live agent that turns an open-ended question into a synthesized, source-aware answer — with its intermediate steps streamed in real time instead of a black-box response.