Multi-Agent Workflow Builder MVP

This project is a template-based Multi-Agent Workflow Builder MVP. The outfit recommendation workflow is now one executable domain example, alongside presentation planning and customer support ticket routing examples.

The current semester scope is intentionally limited:

• Build working workflows across multiple domains.
• Define reusable Workflow Templates.
• Compose custom workflows in the Builder Workspace.
• Generate executable Workflow JSON from templates or workspace inputs.
• Show how the same idea maps to Flowise / Dify / Langflow style builders.

Project Position

Existing tools such as Flowise, Dify, and Langflow already provide general workflow builders. This project does not try to replace them.

Our contribution is the domain-template workflow structure:

User Input
-> Question Node
-> Weather Tool Node
-> Shopping History Analysis Node
-> Recommendation Node
-> Compose Node

In Korean:

기존 Builder를 대체하는 것이 아니라,
개인화 추천 도메인에 필요한 Workflow Template을 정의하고
실행 가능한 MVP로 검증한다.

Main Demo

Run the stable desktop Sample Workflow:

run_desktop.cmd

Example query:

다음 주에 칭다오 여행 가는데 캐주얼 옷 추천해줘

The app runs the 6-node workflow and returns:

• weather-based context
• shopping-history style analysis
• ranked outfit recommendations
• execution trace

Builder Prototype

Run the local Builder Prototype:

run_builder_app.cmd

The Builder Prototype supports:

• reusable node palette
• template selection
• Builder Workspace composition
• required-node validation
• Workflow JSON generation
• generated Workflow execution
• multiple domain workflow templates

Current templates:

configs/builder_templates/outfit_recommendation_template.json
configs/builder_templates/commute_outfit_template.json
configs/builder_templates/presentation_planning_template.json
configs/builder_templates/customer_support_ticket_template.json

Visual Web Verification

If you do not want to inspect backend code, run the visual web page:

run_web.cmd

Then open:

http://127.0.0.1:8000

The page shows:

• selected Workflow Template
• Builder Workspace for composing a custom workflow name and business domain
• Node Palette and sequential Workflow structure
• Builder mapping table
• generated Workflow JSON
• visual run result and execution trace

Harness Engineering Comparison

Run the comparison experiment:

tools\dev\run_harness_comparison.cmd

The experiment compares two construction paths for the same tasks:

• Generic Harness Engineering: manually define Agent roles, tools, context, constraints, verification, and execution order.
• Builder Workspace: select a business preset, name the Workflow, generate executable Workflow JSON, and run the same 6-step trace.

Outputs:

outputs/harness_comparison/harness_comparison_results.json
outputs/harness_comparison/harness_comparison_report_zh.md
outputs/harness_comparison/harness_comparison_report_kr.md

Static Builder Preview

Open this directly in a browser:

workflow_builder_preview.html

This is a static visual explanation page for presentation. The runnable builder is run_builder_app.cmd.

Command Line Verification

List available templates:

python builder_demo.py --list-templates

Show the default builder template:

python builder_demo.py --show-builder

Generate and run the travel/outfit template:

python builder_demo.py --run-generated "칭다오 다음 주 여행 캐주얼 옷 추천해줘" --user user_a

Generate and run the commute/work template:

python builder_demo.py --run-generated --builder-template configs\builder_templates\commute_outfit_template.json "서울 내일 출근 포멀 옷 추천해줘" --user user_b

Generate and run the presentation planning template:

python builder_demo.py --run-generated --builder-template configs\builder_templates\presentation_planning_template.json

Generate and run the customer support ticket template:

python builder_demo.py --run-generated --builder-template configs\builder_templates\customer_support_ticket_template.json

Run project verification:

run_verify.cmd

GPU Baseline Evidence

If the allocated GPU server is still available, run one final stable baseline benchmark for the Local LLM Node evidence:

tools\gpu\run_final_base_benchmark.cmd
tools\gpu\download_final_base_benchmark.cmd

The downloaded files are stored locally under:

_local_artifacts\gpu_results\final_base_benchmark

This benchmark is evidence collection only. It is not required for normal demo execution; the stable demo remains tools\gpu\run_gpu_replay_demo.cmd.

External Builder Tool PoC

These folders document external builder validation paths:

external_tools/flowise_poc/
external_tools/dify_poc/
external_tools/langflow_poc/

Start Flowise PoC entry point:

external_tools\flowise_poc\run_flowise_poc.cmd

Start Langflow PoC entry point:

external_tools\langflow_poc\run_langflow_poc.cmd

Check Dify requirement:

external_tools\dify_poc\run_dify_poc.cmd

Current verified external-tool facts:

• Flowise npm package exists at version 3.1.2.
• Langflow package index shows latest version 1.9.4.
• Dify needs Docker / Docker Compose; Docker is not installed on this machine.
• Flowise local install/start was attempted but timed out because of its large dependency tree, so the stable demo remains the local Python Builder Prototype.

Tests

Run:

tools\dev\run_tests.cmd

Expected result:

28 tests OK

Project Structure

.
|-- configs/
|   |-- outfit_workflow.json
|   |-- outfit_agent.json
|   |-- flowise_poc_mapping.json
|   `-- builder_templates/
|       |-- outfit_recommendation_template.json
|       |-- commute_outfit_template.json
|       |-- presentation_planning_template.json
|       `-- customer_support_ticket_template.json
|-- data/
|   |-- shopping_history.json
|   |-- user_profiles.json
|   |-- presentation_knowledge.json
|   `-- support_policy.json
|-- deliverables/
|   |-- builder_tool_comparison_kr.md
|   |-- competitor_positioning_kr.md
|   |-- dataset_research_kr.md
|   |-- external_builder_tool_verification_zh.md
|   |-- final_acceptance_checklist_zh.md
|   |-- mentor_requirement_alignment_zh.md
|   |-- next_presentation_outline_kr.md
|   |-- ppt_4_slides_builder_addon_kr.md
|   |-- professor_answer_drill_kr.md
|   `-- forms/
|       `-- *.txt
|-- experiments/
|   |-- harness_comparison.py
|   `-- gpu_llm/
|       |-- local_llm_api_server.py
|       |-- run_local_llm_api_server.sh
|       `-- smoke_local_llm_api.py
|-- external_tools/
|   |-- flowise_poc/
|   |-- dify_poc/
|   `-- langflow_poc/
|-- src/
|   |-- agent_builder/
|   |   |-- engine.py
|   |   |-- workflow.py
|   |   |-- template_builder.py
|   |   `-- shopping.py
|   `-- weather_agent/
|-- tests/
|-- tools/
|   |-- dev/
|   |   |-- run_tests.cmd
|   |   `-- run_harness_comparison.cmd
|   `-- gpu/
|       |-- upload_gpu_api_update.cmd
|       |-- start_gpu_local_llm_api.cmd
|       |-- open_gpu_local_llm_tunnel.cmd
|       |-- verify_gpu_remote_node.cmd
|       |-- run_local_llm_remote_demo.cmd
|       `-- run_gpu_replay_demo.cmd
|-- _local_artifacts/       # ignored local packages, GPU outputs, temporary PPTs
|-- desktop_app.py
|-- builder_app.py
|-- web_app.py
|-- builder_demo.py
|-- verify_project.py
|-- workflow_builder_preview.html
|-- run_desktop.cmd
|-- run_builder_app.cmd
|-- run_verify.cmd
`-- run_web.cmd

Current Features

• Sequential 6-node recommendation workflow
• Question Node for missing city/date/purpose/style
• Weather API integration
• simulated shopping history analysis
• ranked recommendation output
• optional Local LLM Node with mock, replay, and remote providers
• local desktop workflow run
• local template-based Builder Prototype
• four templates across recommendation, presentation planning, and support
• generated Workflow JSON execution
• Flowise / Dify / Langflow comparison and PoC entry points
• presentation and Q&A deliverables

Stable GPU Evidence Mode

Use replay mode for the stable GPU-related demo:

tools\gpu\run_gpu_replay_demo.cmd

Replay mode reads a previously downloaded Qwen result JSON and appends it as a Local LLM Node in the workflow trace. It does not require SSH, an open tunnel, or a live GPU API process. The live remote API scripts remain available under tools/gpu/, but they are experimental because they depend on remote process, CUDA, tunnel, and port state.

Out of Scope

The current MVP does not claim to be:

• a full production drag-and-drop Builder
• a complete commercial recommendation engine
• a real shopping-platform OAuth integration
• a local LLM training system
• an enterprise monitoring/admin system