NewsBot.
Autonomous News Research Agent
Research that checks its own work: an agent that searches, evaluates, refines, and synthesizes current information into source-backed briefings.
01 / Overview
Research that checks its own work: an agent that searches, evaluates, refines, and synthesizes current information into source-backed briefings.
02 / Problem
A single web search may return incomplete, outdated, or irrelevant results. A research agent needs to decide whether it has enough evidence before producing a briefing.
03 / Solution
I built a LangGraph workflow that optimizes a query, retrieves web results through Tavily, and evaluates relevance, recency, and sufficiency. Conditional routing returns inadequate results to a refinement loop; sufficient results move to source-backed synthesis.
04 / Architecture
Evaluate → insufficient: refine → search again
Evaluate → sufficient: synthesize → source-backed response
A user asks for a current-information briefing through the application API.
05 / How it works
- Receive a research question through FastAPI.
- Optimize the query and retrieve web results with Tavily Search.
- Evaluate relevance, recency, and sufficiency in LangGraph.
- If evidence is insufficient, refine the query and search again.
- If evidence is sufficient, synthesize a source-backed response.
- Expose reusable search and evaluation tools through an MCP server.
06 / Technology stack
Models a multi-node workflow with state and conditional retry routing.
Supplies web retrieval for current information.
Supports the language-model work in the research and synthesis pipeline.
Exposes the agent through an application backend.
Replaces hardcoded tool bindings with reusable search and evaluation tools.
07 / Engineering challenges
- Evidence quality: evaluate relevance and recency before generating a briefing.
- Incomplete retrieval: conditional routing sends insufficient results through a query-refinement loop.
- Integration coupling: migrating search and evaluation to MCP allows reuse by other agents and clients.
08 / Key features
- Autonomous query optimization
- Real-time web retrieval
- Relevance and recency evaluation
- Conditional retry and self-correction
- Source-backed synthesis
- Reusable MCP tools
09 / Outcomes
The workflow optimizes queries, retrieves web results, evaluates relevance and recency, and refines searches before producing source-backed briefings. Migrating search and evaluation tools to MCP made them reusable by other agents and clients.
10 / Engineering takeaways
Agent behavior is shaped by routing decisions as much as by prompts. Explicit evaluation and refinement steps make evidence quality part of the workflow, while MCP separates tool capabilities from a single agent implementation.