AegisCode

Problem Statement

Autonomous coding assistants often produce hallucinated patches, introduce security vulnerabilities, or overwrite code without proper isolation, test verification, or human sign-off on critical production branches.

Dataset

Multi-repository AST call graphs, commit histories, GitHub issue specifications, and dependency trees parsed in real-time.

Architecture

LangGraph deterministic state machine orchestrating 6 specialized agents (Supervisor, Researcher, Coder, Tester, Security, Reviewer). Backed by FastAPI SSE event broadcasting, MongoDB Atlas audit storage, pluggable Docker sandboxes for untrusted code execution, and a responsive Next.js 15 command center.

Model Selection

Hybrid LLM routing leveraging high-capacity reasoning models for task planning and code architecture, with low-latency models for AST parsing and code reviews. Implements Model Context Protocol (MCP) and Agent-to-Agent (A2A) capability negotiation.

Training Process

Rigorous prompt engineering, bounded repair loops (max 3 diagnostic retry iterations on failed tests), input prompt-injection filtering, and security AST scanning for SSRF, injections, and leaked secrets.

Evaluation Metrics

Results

A production-ready AI software engineering platform deployed across Vercel and Render, capable of autonomous bug fixing, regression testing, and opening production-verified GitHub Pull Requests with full audit trails.

Key Learnings

• 1Multi-agent specialization (separating Planner, Coder, Tester, Security, and Reviewer) prevents single-prompt hallucination cascades.
• 2Running all test executions inside ephemeral, resource-constrained Docker containers is essential for safe autonomous code execution.
• 3Risk-gated human-in-the-loop checkpoints build critical developer trust before mutating production repositories.