Built a real AI agent running on autopilot
We've built a real AI AGENT: @seobotai
It works 100% on autopilot. For example, an article takes 2 hours & runs 727 prompts & 126 jobs. Costs $6.
It's not similar to traditional SaaS.
AI controls a large tree of nested tasks imitating a human org.
How it works (in details) 🧵: pic.twitter.com/pFTNO5bIaw— John Rush (@johnrushx) December 5, 2024
1. User provides a URL of the site to the agent, the work starts here
- it scrapes the URL content and passes it to the next model
- LLM learns all the details about the biz: pricing, audience, solution, problem..
- It creates a full spec and passes it on to the next model…— John Rush (@johnrushx) December 5, 2024
2. Next model creates an SEO strategy
- does keyword analisis
- looks at competitors
- looks at existing articles and pages
- comes up with full knowledge and topic tree
- tests it all for potential and demand via google search console API
- goes to google search to explore more— John Rush (@johnrushx) December 5, 2024
3. Once the strategy is created, the "manager" model creates 100+ tasks.
- then the next level models(100s of them) create 100s of tasks for their "employees/models".
- they execute tasks & report back
- the manager models analyze, give feedback and iterate until they're happy— John Rush (@johnrushx) December 5, 2024
4. The results float up the stream, and every higher level agent can initiate new tasks and ask to improve/redo.
- there are anti-hallucination agents who enter the game at every step to fact-check the outputs. They may initiate more tasks toopic.twitter.com/KZwass67OT— John Rush (@johnrushx) December 5, 2024
5. AI itself isn't that useful without external tools & apis
- some agents work as bridges to the real world. Creative agents ask the bridge agents for work to be done. E.g "go to these sites & put together a list of these items..". (just like a human would ask their assistant )— John Rush (@johnrushx) December 5, 2024
6. When there are so many steps unsupervised by a human, even one mistake gonna break the whole thing
- there are agents who adjust the plan, if some tasks cant be done. e.g. if a given website can't be scraped.
- everything runs on queues & jobs based on "eventuall-consistence"— John Rush (@johnrushx) December 5, 2024
7. The amount of job done by agents is insane. One human would spend days or weeks doing what the agents all together did in 2 hours.
The agent isn't perfect ofc.
Human in the loop improves its output significantly.
We still fight with hallucination and unpredictability of LLMs— John Rush (@johnrushx) December 5, 2024
7. That's it.
Next time, I'll share the architecture of my other AI agent https://t.co/UKFgAlqvII .
It works days and nights looking for new directories and backlinks on internet and then can act in real world, register on websites, submit products, etc.pic.twitter.com/yx3KelMpXY— John Rush (@johnrushx) December 5, 2024
