Multi-Agent Research System with Single-Agent Baseline

What you're building

A research question goes in; a cited brief comes out. The planner decomposes the question into 2–3 subtasks; parallel searchers (web or a provided corpus) gather evidence, one clean context each; the writer integrates findings into a draft; the critic reviews against a rubric with at most one revision loop; a human gate pauses the graph for approve/reject-with-feedback before the final answer. Every handoff is a structured brief, logged to JSONL. Then the part that makes this a portfolio piece: the same 10 questions through a single agent with the same tools, and an honest comparison table.

Suggested structure

# state.py
import operator
from typing import Annotated, TypedDict

class ResearchState(TypedDict):
    question: str
    plan: list[str]
    findings: Annotated[list[str], operator.add]   # parallel-safe
    draft: str
    critique: str
    revision_count: int

# handoff.py — HandoffBrief (pydantic) + log_handoff() -> handoffs.jsonl

# graph.py
# one model for every node AND the baseline — swappable by one string:
# model = init_chat_model("anthropic:claude-sonnet-5")  # or: "openai:gpt-5.5"
builder = StateGraph(ResearchState)
builder.add_node("planner", planner)         # decompose via structured output
builder.add_node("searcher", searcher)       # fanned out per subtask
builder.add_node("writer", writer)
builder.add_node("critic", critic)
builder.add_node("human_gate", human_gate)   # interrupt() lives here

builder.add_edge(START, "planner")
# TODO: fan-out dispatch planner -> N searcher runs (Send-style API)
builder.add_edge("searcher", "writer")
builder.add_edge("writer", "critic")
builder.add_conditional_edges("critic", route_after_critic,
                              {"revise": "writer", "gate": "human_gate"})
builder.add_conditional_edges("human_gate", route_after_gate,
                              {"revise": "writer", "done": END})

graph = builder.compile(checkpointer=sqlite_checkpointer)  # durable!

# baseline.py — single agent, SAME tools, same model, same 10 questions
# compare.py — runs both, sums usage across ALL calls, judges pairwise

Ship it to your portfolio

• README with a 60-second demo — architecture diagram up top, then a short clip or GIF: question in, planner fans out, gate pauses, human approves, cited brief out. Hiring managers look at GitHub before the résumé; make the first minute count.
• Logged inter-agent traces — commit a sample handoffs.jsonl and link it from the README. A multi-agent system with specialized roles and logged inter-agent traces is a specifically-cited portfolio project in hiring guides, and the trace file is the artifact reviewers actually open.
• Demonstrate the HITL approval flow — capture the pause (process exits), then the resume with both an approve and a reject-with-feedback. Approval gates in front of irreversible actions are what enterprises specifically screen for.
• An honest 'Limitations' section — where the single agent won or tied, the judge's noise floor, and what you'd measure next before trusting the multi-agent claim. A measured negative result reads as senior; a suspiciously uniform win reads as rigged.