Hi, I’ve been busy working on two production-grade AI Agent projects, and both are now live online, released on Product Hunt (one reached #36 of the day rank among 200+ products), and actively serving real user traffic.

I’d like to share some of the experiences, technology stack, development process, deployment lessons, and the actual costs involved in building and serving these AI agents. More importantly, I want to show how ideas can quickly become working applications.

First of all, these are not “vibe-coded” rushed projects. Turning ideas into production-ready code, design, deployment, and debugging within one week was still much faster than I expected.

The first project is Craftsman-Agent(https://craftsman-agent.aiagenta2z.com), which turns prompts into buildable 3D assembly charts and step-by-step instructions for creations such as LEGO builds, Minecraft structures, Tesla car wraps, and more.

The platform is designed for creativity and play, targeting:

• game hobbyists
• parents and kids
• 3D design prototyping
• creative builders

Example prompts include:

“How do I build a blue-and-white LEGO yacht?”
“Design a red LEGO F1 race car.”
“Create a Minecraft-style castle with medieval towers.”

The second project is CoachOwl Agent Timetable (https://coachowl.aiagenta2z.com/), an AI Agent timetable, calendar, and orchestration platform that helps users set objectives and assign tasks to both humans and AI agents (such as Codex, Claude, and Gemini) for collaborative work in:

• fitness
• career growth
• learning
• personal development
• productivity planning

The stack I am using:

Server
Python/ Node.js / FastAPI / pnpm serve

LLMs & AI Models
Gemini/ OpenAI /Qwen

APIs(DeepNLP OneKey Gateway)
Google Search APIs
Tavily Search APIs
Nutrition calculation APIs
Gemini Nano Banana image recognition APIs
3D rendering APIs

Deployment & Hosting
AI Agent A2Z platform
One free domain and project per user account (*.aiagenta2z.com)

Coding agent:
Claude
Codex

As you can see, these are fully functioning applications with:

deployed websites and landing pages

• AI Agent tool usage / pipelines
• LLM/Agent Gateway Integration
• External API Fast Integrations (3D/Food Nutrition/Search/Images)

The most important lesson I learned is that shipping production AI systems is far more about orchestration, deployment, debugging, UX, and infrastructure than prompt engineering alone.

1. Craftsman Agent

Design & Development

webpage: https://craftsman-agent.aiagenta2z.com
GitHub: https://github.com/AI-Hub-Admin/Craftsman-agent

I build the main website homepage using claude code/codex in a few minutes. And building the main AI Agent workflow and 3D text-3D instruction using 3 days,
which is more complicated and no free APIs available. The AI Agent workflow of turning text prompts to 3D instruction, for example

text -> 3D parts plans json list -> 3D rendering APIs calling 

Agent Workflow: Text to 3D instruction

For example, a 3D lego boats will needs a inventory of less than 10 to 1000 pieces 3D parts of data in the following formats.
And converting these LLM generated 3D models data into rendering instructions will also involve calling 3D APIs multiple times.
And the raw LLM generated 3D files are actually of pure quality, so SFT and few shots are also needed to enhance the final design.

[
    {
        "color": "red",
        "size": [
            4.0,
            2.0,
            1
        ],
        "position": [
            2,
            1,
            0.5
        ],
        "part_id": 3020,
        "part_name": "Plate 2x4",
        "image": "M3020",
        "categories": "Basic, Architectural, Transportation, Space, Plate, Solid Studs"
    },
    {
        "color": "red",
        "size": [
            2.0,
            1.0,
            1
        ],
        "position": [
            1,
            2,
            1.5
        ],
        "part_id": 3023,
        "part_name": "Plate 1x2",
        "image": "M3023",
        "categories": "Basic, Architectural, Transportation, Space, Plate, Solid Studs"
    },
]

LLM and API Calling: DeepNLP OneKey Agent Gateway of Gemini LLM endpoint and 3D building APIs (https://deepnlp.org/doc/onekey_gateway).

Server/APIs/Skills/MCPs:
The server is developed using python FastAPI backend with endpoints serving both MCPs/Skills and CLIs APIs.
These codes are not open sourced but skills and clis are available to use and registered on OneKey Gateway,
And there are agent run payment paypal endpoint designed to charge per 3D assembly generated.

## endpoint
app = Starlette(
    routes=[
        Route("/chat", chat_endpoint, methods=["POST"]),   # New Chat Endpoint
        Mount("/static", app=StaticFiles(directory= STATIC_DIR.resolve() , html=True), name="static"),
        Mount("/assets", app=StaticFiles(directory= ASSETS_DIR.resolve(), html=True), name="assets"),
        Route("/api/v1/generate_minecraft_build_plan", api_generate_minecraft_build_plan, methods=["POST"]),
        Route("/api/v1/generate_lego_build_plan", api_generate_lego_build_plan, methods=["POST"]),
        Route("/api/v1/generate_tesla_wraps", api_generate_tesla_wraps_build_plan, methods=["POST"]),
        Route("/paypal/agent/purchase/callback", paypal_webhook, methods=["POST"]),
        Mount("/", app=mcp_app),  ## MCP Endpoint Always mounts /mcp
    ],
    lifespan=lifespan,
)

Website Domain & Hosting:
Website: Typescript JS based website serving the /static, /image gallery
https://github.com/aiagenta2z/agent-mcp-deployment-templates/tree/main/quickstart/website_typescript
Doc: https://deepnlp.org/doc/agent_mcp_deployment

Summary:
Cost of time turning the ideas into production AI Agents:

2. CoachOwl AI Agent Orchestration

Features:

• Online Timetable & Calendar: CoachOwl Agent has the calendar & timeline features to allows AI Agent to connect, track, assign and schedule tasks for human and AI Agents to collaborative, competible with Google Calendar, Outlook for AI Agents.
• AI Agent Orchestration: Humans can better assign repetitive & periodic tasks easily to your personal Agents by scheduling, tracking compared to sending messages. Scenario: Repetitively Sending Emails of competitor analysis of ProductHunt daily for 2 weeks. Anaylyzing food calories for 2 weeks. Prepare for SAT/CFA/GRE exams.
• Add a task & objective: You can add task, objective (AI coach will plans several periodic tasks for you and you can always edit task contents.)
• AI Coaches: AI Coaches with Special Skills assigns tasks to both human and agents, such as Fitness Coach, Relationship Coach, Career Coach, Relationship Coach, Fitness Coach: Food Calories Analysis, Image Recognition, Career Coach: Search Indutry info and send briefs to Emails. Prepare for Social Media Anouncement, write Blogs. Learning Coaches
• Easy Voice Input, Habit Tracking, AI Agent Task Scheduling, Connect to Claude Code, Codex, OpenClaw and more.

Support Coaches and Agent Tools & Ability

The app is build on Onekey Agent Router APIs for Image processing, Food Calories Searching, Sending Emails, Deep Research Abilities.

You can always extends more skills to use CoachOwl as an AI Agent Orchestrator to assign tasks to your Local Agents (Codex/Claude/Gemini)

Design & Development

Website Homepage: https://craftsman-agent.aiagenta2z.com

Github(Fully OpenSourced): https://github.com/AI-Hub-Admin/CoachOwl-Agent-Timetable

Deployment: https://github.com/aiagenta2z/agent-mcp-deployment-templates

The AI Agent Orchestration

1. The Timeline (oneday todo list) and the calendar
   The app is build using codex, which costs me one day to finish all the tables design, frontend web design, etc.

2. Agent Execution system
   Agent Execution include setting up agent tasks (recursive tasks), such as "Deep Research the producthunt AI Agent releases in last 7days and send to my Emails at xxxx@gmail.com".
   The agent execution related include
   """
   /claim task -> /Execution -> Report Heartbeat -> Update Results
   """

The db design and implementation costs 4- 5 days. And for local agents running, use "onekey gateway coachowl/coachowl" to pull tasks from the web, which involves
another one day to complete.

1. Summary

Related