Web Research Agent

Web Research Agent

Multi-agent pipeline that takes a user question, breaks it into sub-questions, researches them in parallel, fact-checks the results, and writes a final cited report.

Architecture

User query
    ↓
Orchestrator → breaks query into 3 independent sub-questions
    ↓
Researcher × 3 (parallel, asyncio.gather) → each researches one sub-question via web search
    ↓
Fact-Checker → removes duplication, flags contradictions and implausible claims
    ↓
Writer → synthesizes final report with source citations

Built with the OpenAI Agents SDK, MCP servers for tool access, and structured (Pydantic) outputs at every stage.

Agents

| Agent | Role | MCP servers | |---|---|---| | Orchestrator | Splits query into 3 sub-questions | Sequential Thinking | | Researcher | Answers one sub-question via web search | Sequential Thinking, Web Research | | Fact-Checker | Cross-checks the 3 research outputs | Sequential Thinking | | Writer | Produces final cited report | Sequential Thinking |

Why parallel Researchers

The 3 sub-questions are independent, so running them sequentially would waste time waiting on each one in turn. asyncio.gather runs all 3 at once — total time ≈ slowest single research call, not the sum of all 3.

Structured output

Every agent returns a Pydantic model instead of free text, so each stage can reliably parse the previous stage's output instead of guessing at raw strings:

class SubQuestions(BaseModel):
    questions: list[str]

class ResearchOutput(BaseModel):
    summary: str
    key_findings: list[Finding]
    confidence: float

class Writeroutput(BaseModel):
    final_report: str
    sources: list[Finding]

Setup

pip install -r requirements.txt

example.env file:

baseurl=
api=

config.yaml sets the model, instructions, temperature, and max tokens per agent independently — cheaper/faster models for the Researcher (runs 3× per query), stronger models recommended for the Fact-Checker (the critical accuracy step).

Run:

python main.py

Known limitations

• Source diversity: in testing, all 3 Researcher agents sometimes returned findings from the same single domain, giving the Fact-Checker nothing to actually cross-check. Instructions now ask for 2+ distinct domains, not yet strictly enforced in code.
• Stale data: research results have surfaced outdated (multi-year-old) figures without flagging their age. Instructions now ask agents to note source age explicitly.
• Numeric sanity-checking: early runs returned an implausible price-per-watt figure that passed through unflagged. Fact-Checker instructions now ask it to flag numbers that look logically inconsistent.
• No Critic/review loop: the Writer's output is not currently re-checked for citation completeness before being shown to the user.
• MCP on Windows: Sequential Thinking / Web Research MCP servers have shown occasional timeout issues on Windows.

Next steps

• Add a Critic agent to review the Writer's output and force a re-write if citations are missing
• Enforce domain diversity across Researcher calls in code, not just via prompt instructions
• Wrap in a FastAPI endpoint for external use

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: MilanSirko
• Source: MilanSirko/Web-Research-Agent
• License: MIT

Install and usage instructions live in the source repository linked above.