English
AI Agents Review: ChatGPT, Genspark, Replit and Manus Tested on Real Business Projects
The Best Way to Rest Is to Work — and Let Others Work for You
What is the best way to rest? Rest and work at the same time.
Even better — work in a way where others do the work for you. Like AI agents.
Over the holidays I ran a full test drive of different neural networks and AI agents.
Not just “playing around” — but on real products that are either already live or currently in demand-validation stage.
Below is a breakdown of my experience with each tool and the products built using it. The projects: an SMM manager for my clubs, a tool rental and decorative materials sales business (chosen specifically to launch fast in a totally unfamiliar space), and a personal assistant.
This post was written with the help of yet another agent that tracked everything and took notes.
Hope it’s useful. If it was, drop a fire — I’ll share the rest of the story.
AI Agents Review: Starting with ChatGPT as the Architect and Quality Control Layer
ChatGPT served as the base layer across all projects.
I used it to:
- conduct deep market research (rentals, decorative materials, content products),
- understand the economics of new businesses,
- write technical specifications for services and Telegram bots,
- build checklists for validating agent performance,
- use it as a companion assistant and “second brain.”
Specific deliverables:
- Spec for a Telegram digest from 15 channels in audio format
- Full spec and training logic for an SMM manager for the Zornitsa and Expedition clubs — and, critically, step-by-step oversight of the bot’s training process plus a third-opinion layer
- Architecture for a Telegram bot for a tool rental service
- Ad copy and category structure for Avito listings (flexible stone and brick). Again, critically — step-by-step control over ad uploads to maximize Avito’s algorithm performance
- Validation of DEV/PRODUCTION logic and cloud storage setups. This one deserves its own story. Many vibe coding tools tell people like me — who are far from programming — absolutely nothing about how dev and production environments work. You end up burning a ton of credits before you figure it out. ChatGPT walks you through it step by step.
In the end, ChatGPT is not an executor — it is an architect and a controller.
It rarely produces the final artifact, but it constantly keeps the system from breaking. It doesn’t handle tables or images well, and it can’t even format complex text properly. But as an assistant and a third eye — it’s superb.
AI Agents in Action: Genspark as the Primary Executor
Genspark was my main execution agent.
This is where the actual work happened:
- Telegram digest (channel aggregation, audio generation, send scheduling),
- SMM manager for the Zarnitsa and Expedition clubs,
- visuals and images for decorative materials,
- product catalogs and presentations,
- initial site assembly and admin panels.
Specific deliverables:
- Voice Telegram digest from 15 channels
- SMM manager for the clubs
- Catalog and presentation for decorative materials
- Image generation via NanoBanana integration
Genspark is excellent at launching and packaging —
but struggles with architecture:
- confuses DEV and PRODUCTION environments,
- loses context,
- freezes,
- forgets scheduled sends for automated messages (the audio digest bot).
Even the inspector agent I set up to monitor the main bot started glitching over time.
That said — it builds presentations beautifully, and it is genuinely the first service (after trying about a dozen) that actually satisfied me in that area. Image generation is great. Handles Excel spreadsheets at a solid B level. Builds landing pages fast and grasps the idea instantly.
Replit: The Stable Foundation for Production
Replit became the stable base when the product simply needed to stay alive.
Specific deliverables:
- Not just a website, but a full catalog — a site and admin panel for the decorative materials business
- Product card storage
- The production environment where Genspark and Manus pushed data
Replit:
- is not the “smartest” tool,
- but it holds the structure together,
- doesn’t mix up environments,
- and stays stable in production.
Manus: Scalable Hands for High-Volume Tasks
Manus was used where the job required manual-style work at scale.
Specific deliverables:
- Competitor site analysis (decorative materials).