Open-source agent-engineer plugin boosts AI coding assistant requirement compliance to 53%

Sarthak Rastogi released agent-engineer, an open-source plugin designed to make coding assistants behave like senior AI engineers. The plugin supports Claude Code, Codex CLI, Cursor, OpenCode, and Gemini CLI, featuring 11 skills, 4 subagents, and 5 hooks without collecting data or requiring a backend.

Performance Validation

A blind A/B test using Claude Code (Opus 5.5) across six agent-engineering tasks demonstrated a significant improvement in quality. With the plugin applied, the requirement fulfillment rate reached 53%, compared to 34% without it. The evaluation used pre-written checklists and three runs per task.

Core Architecture and Skills

The plugin addresses common agent failures such as false positives from simple tests, symptom-only patches, and rule drift in long sessions. It employs a Router Skill that fixes issues at the lowest explanatory layer, moving to complex structures only when evidence supports it.

Key skills include:

• Design: Prioritizes simple workflows over autonomous agents and requires success criteria before architecture.
• Tools: Enforces clear names, typed schemas, and bounded response sizes for better model understanding.
• Evals: Mandates trace-based error analysis and CI gate integration, reducing unmeasured claims from 3/3 to 1/3 in tests.
• Observability: Generates spans for LLM and tool calls using OpenTelemetry GenAI conventions, treating prompts as PII.
• Guardrails: Defends against jailbreaks and prompt injection, optionally integrating the rival-ai Python library for harmful query classification.

Installation and Usage

Users can install the plugin via git clone or npx skills add sarthakrastogi/agent-engineer. It operates without explicit commands, loading appropriate skills based on natural language requests. The plugin stores long-term decisions in an agent-engineering/ folder to maintain context across sessions.