KISS Sorcar: A Stupidly-Simple General-Purpose and Software Engineering AI Assistant

2026-06-30 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

KISS Sorcar is a general-purpose and software engineering AI assistant implemented as a free, open-source Visual Studio Code extension. Built on the 1,850-line KISS Agent Framework, it addresses common LLM agent limitations like finite context windows, session derailment, and "AI slop." The system employs a robust prompt and a five-layer hierarchy, enabling budget-tracked ReAct execution, automatic task continuation via summarization, coding and browser tools, persistent multi-turn chat, and git worktree isolation for task-specific branches. Prioritizing output quality over speed, KISS Sorcar validates its work with linters and tests. It achieved a 62.2% pass rate on Terminal Bench 2.0 using Claude Opus 4.6, outperforming Claude Code (58%) and Cursor Composer 2 (61.7%).

Key takeaway

For AI Engineers deploying LLM agents for software development, KISS Sorcar demonstrates that a layered, single-concern architecture combined with rigorous prompt engineering can significantly improve agent reliability and performance. You should consider adopting similar principles, such as explicit budget tracking, automatic task continuation, and git worktree isolation, to mitigate common LLM agent failures and enhance long-horizon task completion. This approach prioritizes output quality, reducing "AI slop" and increasing confidence in agent-generated code.

Key insights

Layered architecture and disciplined prompting significantly enhance LLM agent reliability and performance in software engineering.

Principles

• Adopt a "Keep It Simple, Stupid" design with isolated concerns.
• Prioritize output quality over latency for robust agent performance.
• Mandate a test-first discipline for all bug fixes and code changes.

Method

A five-layer agent hierarchy manages budget-tracked ReAct execution, automatic task continuation via structured summarization, coding/browser tools, persistent multi-turn chat, and git worktree isolation.

In practice

• Utilize `USER_PREFS.md` for agent self-improvement and preference learning across sessions.
• Implement real-time budget tracking to ensure cost transparency and control.
• Integrate browser automation for in-IDE web interaction and verification.

Topics

• AI Agents
• Software Engineering
• Prompt Engineering
• LLM Architecture
• Git Worktrees
• VS Code Extension

Code references

• ksenxx/kiss_ai
• harbor-framework/harbor
• algorithmicsuperintelligence/openevolve
• paul-gauthier/aider
• gsd-build/get-shit-done

Best for: Research Scientist, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• 4 Layer "AI Harness" For LLMs (+54%). Really? — The "AI Harness" improves LLM agent performance by addressing interface mismatches and syntax issues, not by modifying the frozen model.
• I Stopped Writing Longer Prompts and Started Writing “Skills.” — Encapsulating engineering procedures into contextually loaded "skills" dramatically improves AI agent reliability and discipline.
• Every Software Just Became Agent-Native — AI agents are gaining universal software control and self-improvement capabilities through CLI harnesses and skill optimization.
• Fragments: April 21 — AI's rapid advancement necessitates revisiting foundational software principles and developing robust security for "permission hungry" agents.
• Clever Prompts Are Cheap Now. Reliable LLM Prompting Systems Are the Skill. — Reliable LLM systems require engineering discipline, moving beyond single clever prompts to structured, validated, multi-step workflows.
• I Applied Andrej Karpathy’s Auto-research to Software Development — Integrating LLM agents with code-defined invariants enhances autonomous software development iteration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.