Skip to content

tage64/CoderAgent

Repository files navigation

Coder Agent

This is an experimental project to create a coder agent which uses LLMs to write Python code.

Procedure

  1. Ask the user for a problem specification.
  2. Use the "programmer agent" to write one or more python functions which solves the problem. It is important that the functions doesn't perform any side effects like printing or writing to the file system as that is hard to test. The system prompt for the agent includes instructions to add type annotations and write docstrings for all functions.
  3. Type checking is performed in the following steps: a) The code is run through the static type checker Mypy. If it passes type checking, then we are done, else goto step b). b) The programmer agent is passed a prompt with the original code, the problem specification from the user, and the type errors, it is asked to solve the type errors. c) Go back to step a) if not the maximum number of retries is reached, then exit.
  4. Ask the "test designer agent" to write tests for the functions: a) Generate a so called stub file with Stubgen. A stub file is basically all function definitions with type annotations and docstrings but excluding function bodies. b) Feed the problem specification and the stub file to the test designer agent and ask it to write tests for all functions using unittest.
  5. Merge the code and the tests into the same file and type check in the same way as in step 3.
  6. Run the tests. If they succeed, we're done, else go to step 7.
  7. Feed the code (but not the tests) and the test errors to the programmer agent and ask it to update the code so that it passes the tests.
  8. Go back to step 5 and retry with the updated code. There is of course a maximum number of retries here as well.

Running

You need at least Python 3.10 and Python Poetry installed. You can then run the following commands:

poetry install
poetry run coder_agent

You may specify arguments like --backend, --temperature, --retries and --no-interactive on the command line. For reference run:

poetry run coder_agent --help

Example Queries

  • Write a function which sums two numbers.
  • Write a function which computes the factorial of an integer.
  • Write a function which computes the squared factorial of an integer.

Limitations

One limitation is that the tests will never be improved. So if there is a bug in the tests the programmer agent has no chance of correcting the code.

About

An experimental coder agent based on an LLM

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published