Skip to content
NNadeem Shabir
Back

Project

Autonomous SWE Agent

A coding agent that reads the repo, edits, and looks at the tests

Given a coding task, it retrieves repository context, edits files, runs tests, and records the run. A failed tool call comes back as an observation instead of ending the job.

  • Python
  • FastAPI
  • Tree-sitter
  • ChromaDB
  • Celery
  • PostgreSQL
01

What it does

Repository work is more than a code suggestion. The agent has to find the relevant files, change the right places, run tests, and respond when an assumption is wrong.

02

Retrieval

  • Tree-sitter parses source into chunks. Sentence Transformers embed them. ChromaDB stores the index.
  • The agent pulls the code that matters for the task, instead of dumping the whole repository into the prompt.
03

Loop

  • A manual ReAct loop alternates between a decision and a tool: search, read, edit, run tests, read the output.
  • Tool results go back into the next decision. A tool error is an observation. It does not have to stop the run.
  • Planner, coder, and reviewer are stages in that one loop, not three separate services.
04

Limits and backend

  • An execution budget caps how much a run can do. File edits stay inside the working tree.
  • FastAPI accepts the request. Celery runs the agent. Redis is the queue. PostgreSQL stores the run and the trace. Docker Compose hosts the stack on an Azure VM.

Tested on controlled GitHub issues in a test repository. A wide range of real repositories still needs further evaluation.