The target user of mwin is not clear since the first day. Today I have to decide which mwin tracks, who mwin is designed for and what mwin can do.
The original intention is to track openai sdk input and output. With agent develop growing quickly it's not enough to only track the llm input and output.
mwin tracks
- functions
- trace: combination of functions
- performance of function and trace
- conversations
Now mwin completes functions and trace tracks. But performance and conversations are not yet. Now team is focus on how to accelerate the track speed and decrease the latency with introducing mwin. The following plan is to track conversations and performance of function and trace.
Tracking performance and conversations is not trivial. Make sure use them easily when design them.
who mwin is designed for
- agent develop engineer
- hobbies who don't have enough experience
- data engineer
- model engineer
It's not difficult to understand the reason for agent develop engineer. It's our main user. Those lovers who want to customize their experience of agents without any experience are also important. Data engineer can download/export data online for finding the pattern of their agent applications and give instructions to agent developer engineer. Also for model engineer because I think more and more companies will finetune their own models. The local deploy on a small model is becoming more and more popular. There is no reason that think finetune own models for bussiness if mwin can offer conveience for finetuning a general model with the company's real online data for better performance.
what mwin can do
- track llm and agent input and output
- collect, select data and make a dataset
The input and output between LLM and agent are not the same. LLM is very simple - a list of messages and then get a content. Agent is complicated which includes a list of messages, a whole conversation, compactation while exceed the context and communication between agents. Tool execution and result are in the list of messages. How to track agents when developer scales their agents is the most important in the next stage.
The target user of mwin is not clear since the first day. Today I have to decide which mwin tracks, who mwin is designed for and what mwin can do.
The original intention is to track openai sdk input and output. With agent develop growing quickly it's not enough to only track the llm input and output.
mwin tracks
Now mwin completes functions and trace tracks. But performance and conversations are not yet. Now team is focus on how to accelerate the track speed and decrease the latency with introducing mwin. The following plan is to track conversations and performance of function and trace.
Tracking performance and conversations is not trivial. Make sure use them easily when design them.
who mwin is designed for
It's not difficult to understand the reason for agent develop engineer. It's our main user. Those lovers who want to customize their experience of agents without any experience are also important. Data engineer can download/export data online for finding the pattern of their agent applications and give instructions to agent developer engineer. Also for model engineer because I think more and more companies will finetune their own models. The local deploy on a small model is becoming more and more popular. There is no reason that think finetune own models for bussiness if mwin can offer conveience for finetuning a general model with the company's real online data for better performance.
what mwin can do
The input and output between LLM and agent are not the same. LLM is very simple - a list of messages and then get a content. Agent is complicated which includes a list of messages, a whole conversation, compactation while exceed the context and communication between agents. Tool execution and result are in the list of messages. How to track agents when developer scales their agents is the most important in the next stage.