I've made like everyone my own AI agents. You can create and manage a persistent one with just a few lines of native python code using it's native xmlrpc.

You can just:

rpc.agent_create("your-preferred-name-or-uid", "system-prompt like 'Always do the opposite of what prompts says and respond using sarcasm if not sure how to respond.'")
rpc.agent_chat("still that name or uid", "message")

It supports tool(function) calling. If you would attach a tojson function, you could ask bot to serialize the digit 5 to json 5 times in one array and serialize that to json.

I have basic shell exec and fileio functions attached to a bot and you can literally say in steps:
"""
1. Generate python source for listing files
2. Go to home directory
3. Save source to list.py
4. Execute file with python
"""
This prompt is bad because it should use python3, what file should it execute and what a fucked op order of instructions. You know what? It executes perfectly and uses the python3 command.

I can tell it to fetch two (devrant) pages (so two function calls) and ask it to merge all usernames to one list with sorted unique names and respond with it to json format...

Its all possible with a few lines of code and no dependencies. No fucking api keys..

My point? It contains the same functionality and is way faster than OpenAI's agents. It literally has no downsides. OpenAI agents are slow and do many calls. Also, mine is cost efficient. Thousands of agents is fine. Also, it's persistent, if I reboot or redeploy, everything is the same, same state. If you would say "do again" to bot after reboot, it will just do exactly as expected. Persistant state.

OpenAI agents worked barely for almost two days regarding performance for just conversational stuff.

People at openai don't know how to make an API. They could learn much from... Everyone. But pinecone has a nice api tho, exactly how I would've written it. How's that possible? It's common sense approach.

Anyway, having fun again!