Finance Demo

Multi-Agent Ticket Review

A review pipeline where parser, quality inspector and knowledge matcher collaborate through structured state.

Runnable demo · Repo path: workshops/workshop2-ticket-review

Structure Blueprint

Clickable pattern simulator

Interactive Browser Demo

This simulator does not replace the CLI run, but it makes each step, state, log, gate and final output visible in the browser.

Live pipeline

State inspector

Execution log

Current output

Step-by-step Design

Raw Ticket

The input remains a messy operational object: text, priority, component, environment, and user signal.

Context Triage

Parser Agent

The parser converts prose into structured JSON so the rest of the graph can reason on fields.

Tool Dispatch

Quality Agent

The quality role checks reproduction steps, environment, expected behavior, and evidence completeness.

Adversarial Review

Knowledge Agent

The knowledge role searches related tickets and detects duplicates or known incidents.

RAG Pipeline

Merge Node

Parallel findings are gathered into one decision state with score, risks, and routing reason.

Fan-out/Gather

Human Review

Weak, duplicate, or high-priority cases are escalated instead of auto-closed.

Approval Gate

Screenshots

Three sample tickets: approved, needs human review, potential duplicate.

Pattern Mapping

Pattern	Design role
Handoff Chain	Each agent owns one responsibility and passes a structured report.
Fan-out/Gather	Quality and knowledge checks can run in parallel and merge later.
Adversarial Review	A second role challenges ticket completeness and priority claims.
RAG Pipeline	Historical ticket knowledge is retrieved as evidence, not guessed.
Approval Gate	Risk tier determines auto-pass, revision, or human review.

Design Insights

Topology before agents

Do not add agents first; decide whether work should be sequential, parallel, or review-gated.

JSON is the interface

The key collaboration surface is structured state, not free-form conversation.

Risk routing is the product

The system is useful because it decides what can move and what needs attention.

Run Path

cd workshops/workshop2-ticket-review
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cd src
python main.py

# optional visual walkthrough
open ../visual/index.html

READMEworkshops/workshop2-ticket-review/README.md

Blueprintimages/training/visual/workshop2-pattern-blueprint.png