ThesisResearchAI/ML 8-12 months
Autonomous Multi-Agent System for Software Development
A multi-agent LLM system where specialized AI agents collaborate to autonomously develop software — Architect agent designs the system, Coder agent writes code, Reviewer agent critiques, Tester agent validates.
PythonLangGraphOpenAI/Anthropic APIDocker SandboxVector DB (Chroma)GitHub API
Find GitHub repos & source code →How to build it — step by step
- 1Agent Design: Define agent roles with specific system prompts and tool access: Architect (diagrams), Coder (file writes), Reviewer (static analysis), Tester (test execution)
- 2Orchestration: Use LangGraph state machine to manage agent handoffs; shared workspace via file system and vector DB
- 3Code Execution: Sandboxed Docker containers for safe code execution; capture stdout/stderr for feedback loop
- 4Evaluation: Benchmark on SWE-bench: measure patch correctness, test pass rate, and code quality metrics
Key features to implement
- ✓Fully autonomous code generation pipeline
- ✓Agents communicate via structured messages
- ✓Iterative refinement until tests pass
- ✓Human-in-the-loop approval gate for deployment
- ✓Benchmark evaluation against human developers
💡 Unique twist to stand out
Study agent communication failures and propose a "Consensus Protocol" where agents vote on design decisions with a mediator LLM resolving disagreements — analyzing how disagreement resolution affects final code quality.
🎓 What you'll learn
Multi-agent AI architecture, LLM orchestration frameworks, autonomous software engineering, and rigorous AI systems evaluation methodology.