Skip to content

Nexus

Nexus is an autonomous AI engineering team that turns one instruction into verified code patches, eliminating coordination busywork between PMs, frontend, and backend engineers.

Code Video Cerebral Valley

Core trick

Six role-specific agents produce unified diffs that are applied with git and verified with a real build before the run is accepted.

Tags

dev toolsagents

Description

Nexus is an autonomous AI engineering team that turns one instruction into verified code patches, eliminating coordination busywork between PMs, frontend, and backend engineers. Six Claude Opus 4.6 agents work together: Router decides focus, Frontend and Backend agents scan with domain expertise, Manager prioritizes like a PM, Detect reviews code deeply, Plan creates tasks, and Patch generates real unified diffs applied via git and verified with a real build. Opus 4.6 powers six distinct steps with different system prompts, extended thinking visible in the UI, and structured JSON contracts between agents. A Reflector saves learnings to persistent memory so every run is smarter, and the system also learns from external articles automatically. No hardcoded rules, every decision is Claude. The demo shows three flows on a real Next.js codebase: scan mode, instruction mode with auto-apply and build verification, and continuous learning from the web.
rashmi-star/claude-hackathon★ 10 forksTypeScriptCreated Feb 16, 2026 · last push Feb 16, 2026

README excerpt

# Nexus — Autonomous AI Engineering Team

**One instruction in, real patches out.** Built entirely with **Claude Opus 4.6**.

Nexus is an autonomous AI engineering team that turns a single natural-language instruction into verified code patches, eliminating the coordination busywork between PMs, frontend, and backend engineers. Six Claude Opus 4.6 agents work as a coordinated team: Router decides the focus area, Frontend and Backend agents scan with domain expertise, Manager prioritizes and assigns like a PM, Detect performs deep code review, Plan creates targeted tasks, and Patch generates real unified diffs, applied via git and verified with a real build. Opus 4.6 isn't used as a single prompt. It powers six distinct steps with different system prompts, extended thinking visible in the UI, and structured JSON contracts between agents. A Reflector agent saves learnings to persistent memory so every run makes the next one smarter, and the system also learns from external technical articles automatically. No hardcoded rules. Every decision is made by Claude.

---

## How It Works

```
Instruction
    ↓
[Claude: Router]        ← decides: frontend, backend, or full-stack
    ↓
[Frontend Agent]  [Backend Agent]   ← Claude scans the repo with domain expertise
    ↓
[Claude: Manager]       ← prioritizes issues, assigns to the right agent
    ↓
[Claude: Detect]        ← deep code review with memory from past runs
    ↓
[Claude: Plan]          ← creates 2-3 targeted tasks
    ↓
[Clau

Team