SAGE IDE — Adversarial Multi-Agent Code Engine

SAGE IDE is a browser-native AI coding environment built on one radical premise: code quality improves when AI systems argue. Unlike single-agent assistants, SAGE deploys four specialized agents into a live adversarial boardroom. The Architect designs. The Implementer builds. The Synthesizer integrates. The Red-Team attacks — mid-stream, without waiting for a review phase. A challenge bar fires in real time, quotes the vulnerable excerpt, and assigns a severity tag (Critical/High/Medium/Low) with CVE reference. The Nash Crucible then aggregates agent votes, detects equilibrium, and locks the final output as a signed, immutable Artifact. The adversarial layer is the core innovation, not an added feature. Every agent has a permanent seat and the authority to interrupt others mid-stream. The boardroom is the interface — not a sidebar, not a chat panel. Agents generate structured Artifacts (Design Plans, Diffs, Challenge Reports, Locked Outputs) that are reviewable, commentable, and auditable. Built on AMD Instinct MI300X via ROCm: four co-resident Qwen/Llama models sharing 184 GB VRAM through vLLM, locked into dedicated memory regions per agent. Backend: FastAPI + LangGraph (7-node graph) + Celery + Redis for async agent parallelism. Frontend: React + Monaco Editor with live WebSocket streaming and gutter attribution marks per agent. CPU fallback (sage-free) validated for non-GPU environments. Benchmarks: +12.4% HumanEval+, +8.2% SWE-Bench-Lite over single-agent baseline. Better process, not a bigger model. What ships: LangGraph multi-agent engine, WebSocket layer, React IDE with Monaco, Agent Boardroom with live challenge bars, Nash Crucible, and full Artifact system — demonstrated end-to-end, no mocks. GitHub: github.com/realruneett/Sage