By admin

30 thoughts on “Introducing Micro Agent: An (Actually Reliable) AI Coding Agent”
  1. But then the problems it fixes still exist in the test itself. The tests are more likely to be full of errors themselves. Then what? Ask it to make a test to the test, infinitely? Fix the test yourself? That'd go against the whole point of the agent.

  2. steve, ive been following all this ai videos and your approach of instead having big models, to create mini/micro/small models/agents to solve "one" problem. great work, man! you are inspiring

  3. Interesting.

    I have for some time held the belief that to improve LLM coding you would need something like this. Good to see you working on this.

    Beyond "tester" and "coder" personas I'd like security and software quality personas as agents also.

  4. By YouSum Live

    00:00:00 Frustration with traditional AI code generation
    00:00:32 Proposal of a micro agent for code generation
    00:00:40 Micro agent converts prompts into tests first
    00:00:57 Iterative code development based on test results
    00:01:06 Clear success criteria through test-driven development
    00:01:18 Feedback loop enhances AI's debugging capabilities
    00:02:26 Integration of microagent into development workflow
    00:02:45 Future applications of microagent in diverse use cases
    00:03:05 Continuous improvement and feedback encouraged

    By YouSum Live

  5. I tried making the same thing in December last year. But the problem I kept running into is that the tests themselves are wrong af. I soon gave up on the project

  6. I know it's just an example, but it implemented groupAnagrams in O(n * mlogm) where n = strs.length and m is avg size of a given string when in fact you can do all that all in O(n * m) without the sorting.

    And therein lies the problem with AI code slop. Imagine if the code implementation was much longer and complex; these things become harder to spot, especially once you get used to them. Don't get me wrong though, I still think this approach is very very useful in other domains of work (Figma, prototyping, etc). I just wouldn't use micro agents for code gen in any real setting.

  7. The entire flow living inside command line might be somewhat annoying

    The quality of the final output depends on the quality of the tests

    If we could edit the tests in vs code and then just have a command that gets a test file and outputs the implementation that would be great 🙂

Leave a Reply

Your email address will not be published. Required fields are marked *