p-ferreira / agent-subnet Goto Github PK

Here are the tasks:

1. {'description':'Find libraries', 'acceptance_criteria':'A suitable list of libraries is found wiht well-justified explanations','dependencies':()}
2. {'description':'Write main.py', 'acceptance_criteria':'main.py contains functions a b c which do x,y,z','dependencies':(1)}
3. {'description':'Write test.py', 'acceptance_criteria':'tests all functions in main.py','dependencies':(1,2)}
4. {'description':'Demo app', 'acceptance_criteria':'Run main.py and show that it prints hello world, and that test passes','dependencies':(1,2,3)}
5. {'description':'Design logo', 'acceptance_criteria':'Logo effectively communicates the brand','dependencies':()}

Uncompiled (scores source code)

• agents scoring code with LLM
• code health checks
• security checks

Compiled (scores application)

• agents scoring and running code locally
• profiling
• benchmarking
• unit tests

project plan -> tasks -> functions | Write doc string -> generate function

MetaGPT

We want to adapt the structure of MetaGPT to be a more suitable fit for a subnet

Validator

Validator will be the project manager

• Generates requirements
  • Makes a project plan
  • Creates tasks
• Decides who to query for each task
• Evaluates each task objectively

Miner

Miner can be any of the following roles:

• Software engineer (stage 1)
• Architect (stage 2)
• QA (stage 3)

Querying

For MVP we will keep this simple - miner is just a SE. For this reason we don't need any intelligent routing of queries based on miner roles. We can select any UID(s) for any step.

The other question is whether the same miner builds the entire product or a miner just contributes a single step.

For now, a random subset of miners will be assigned a task during each step.

If our project manager can also create a dependency attribute for each task we can see which tasks can be executed simultaneously. This is cool because then in a single step we can perform multiple tasks.

Evaluation

We will make use of objective measurements to determine the quality of miner completions #1 . This is preferred to a QA agent which still may be wrong about code quality.

It may be easier and more robust if we rely on a github repo as the codebase rather than constantly sharing code snippets and sending over the network #3