KISS Sorcar: A Stupidly-Simple General-Purpose and Software Engineering AI Assistant
KISS Sorcar is a free, open-source VS Code extension built on a minimal 1,850-line agent framework that employs a five-layer hierarchy, self-validation via testing, and git isolation to achieve a 62.2% pass rate on Terminal Bench 2.0, outperforming existing tools like Claude Code and Cursor Composer in long-horizon software engineering tasks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a brilliant, hyper-intelligent intern named Sorcar. This intern is incredibly talented at writing code, solving puzzles, and using computer tools. However, like any intern, they have a few quirks: they get tired easily (running out of "brain space"), they sometimes make a small mistake that ruins their whole day, and they can get stuck in a loop of confusion.
The paper introduces KISS Sorcar, a system designed to manage this intern so they can do serious software engineering work without falling apart. The name "KISS" stands for "Keep It Simple, Stupid," which is the core philosophy: instead of building a massive, complicated robot brain, the creators built a very simple, layered management system.
Here is how the system works, explained through simple analogies:
1. The Five-Layer Management Team
Instead of one giant brain trying to do everything at once, the system uses a team of five managers, each with one specific job. Think of it like a relay race where the baton is passed up a chain of command:
- Layer 1: The Runner (KISS Agent)
This is the basic worker. It follows a simple rule: "Think, then do, then think again." It has a strict budget (like a wallet) and stops if it runs out of money or steps. It's the engine that actually talks to the AI model. - Layer 2: The Note-Taker (Relentless Agent)
The Runner has a short memory (a small notepad). If a task is too big to fit on one page, the Note-Taker steps in. When the Runner fills the page, the Note-Taker writes a detailed summary of what happened so far, clears the page, and starts a fresh session. This allows the intern to work on a project for days without forgetting the beginning. - Layer 3: The Tool-User (Sorcar Agent)
This layer gives the intern a toolbox. It can open files, type commands, browse the internet, and even control a web browser. It's the "hands" of the system. - Layer 4: The Chat-Partner (Chat Sorcar Agent)
This layer remembers the conversation. If you talk to the intern today and come back tomorrow, this layer ensures the intern remembers what you discussed yesterday, so you don't have to start over. - Layer 5: The Safety Guard (Worktree Sorcar Agent)
This is the most important safety feature. Imagine the intern is working on a painting. Instead of painting directly on your original canvas, this layer gives them a separate, identical copy (a "worktree") to practice on.- If they mess up, they just throw away the copy. Your original painting is safe.
- If they do a good job, you can carefully copy just their changes onto your original canvas.
- This prevents the intern from accidentally breaking your entire codebase.
2. The "Strict Teacher" (The System Prompt)
The paper argues that the AI model itself is smart, but it needs strict rules to behave like a professional engineer. The creators wrote a massive "rulebook" (the System Prompt) that acts like a strict teacher.
Instead of just saying "Write code," the rules say:
- "Read before you write": Never touch a file until you've read it first.
- "Test before you fix": If you find a bug, write a test to prove it's broken, then fix it.
- "No AI slop": Don't write fluffy, useless text. Be direct and accurate.
- "Check your work": Before you say "I'm done," you must run a checklist (linters, tests, type checks) to prove it works.
The paper claims that by forcing the AI to slow down and follow these engineering habits, the quality of the code goes up dramatically, even if it takes a little longer.
3. The Results: A Magic Trick?
The creators tested KISS Sorcar on a difficult benchmark called Terminal Bench 2.0, which is like a series of complex coding puzzles.
- The Score: KISS Sorcar solved 62.2% of the tasks.
- The Competition: It beat other famous AI coding tools (Claude Code at 58% and Cursor Composer 2 at 61.7%) using the exact same underlying AI model.
The paper claims this proves that organization and discipline matter more than just having a "smarter" model. A simple, well-managed system can outperform a more complex one.
4. The "Self-Teaching" Feature
One cool feature is that the system learns from you. If you tell the intern, "I hate using camelCase for variable names," the system writes this down in a special file (USER_PREFS.md). Next time you ask for help, the intern remembers this rule automatically. It's like the intern is slowly becoming a perfect clone of your personal coding style.
5. The Big Conclusion
The paper ends with a bold observation: We are entering a "Golden Age of Experts."
- If you are an expert who knows how to talk to these advanced AI models (like Claude Opus 4.6), you can build incredible things.
- The AI is so smart that the human might actually be the "slow part" of the process.
- The authors suggest that in the future, people will stop relying on giant software companies and instead build their own custom software tools using these AI assistants.
In a nutshell: KISS Sorcar is a simple, open-source tool that wraps a powerful AI in a strict set of engineering rules and safety nets. It treats the AI not as a magic wand, but as a highly skilled worker that needs a good manager, a clean workspace, and a checklist to do its best work.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.