Agentic-Dev on ROBOCO

Cognitive Debt - Managing a New Kind of Debt in the Vibe Coding Era

Sat, 18 Jul 2026 10:00:00 +0900

“Shipping first time code is like going into debt. A little debt speeds development so long as it is paid back promptly.” — Ward Cunningham, OOPSLA ‘92¹

Dohyun Jung - Principal Consultant, ROBOCO

Work with vibe coding long enough and a strange moment arrives. This is unmistakably my project, yet when I open the code it feels like someone else’s house. The agent designed it, the agent implemented it, and I pressed the approve button. The work got done fast. But when something breaks, I find myself saying: “Why did this happen? The AI wrote it, so I’m not really sure.”

Does Karpathy's LLM Wiki Actually Work? A 72-Run Benchmark

Thu, 07 May 2026 11:00:00 +0900

“To keep agents from re-deriving everything every time, you need a compiled knowledge artifact.” — Andrej Karpathy, LLM Wiki (GitHub Gist, 2026-04)¹

Dohyun Jung - Principal Consultant, ROBOCO

TL;DR

We measured whether Andrej Karpathy’s LLM Wiki pattern (“don’t re-read every time; reuse knowledge compiled once”¹) is actually a win, using an 8-task × 3-technique × 3-repetition = 72-run benchmark in a 30-repository migration workspace.
LLM Wiki placed first on all three dimensions: tokens, time, and quality. Versus Vanilla, 54% fewer tokens and 39% less time (statistically a large effect, §4), with quality equal to Vanilla and better than Graphify.
Graphify (GraphRAG) is unsuitable as a default tool. Negligible token savings, the longest time, and the lowest quality.
That said, it’s not universal. Wiki wins decisively on tasks that synthesize multiple sources, but for tasks whose answer sits plainly in a single source (exhaustive grep, looking up a specific table), reading directly is faster and more accurate (§3).
On receiving these results, the project’s default context-retrieval tool was switched to LLM Wiki that same evening.

Technique	Tokens (avg)	Time (avg)	Quality (out of 25)
Vanilla (direct read/grep)	755K	167s	15.1
LLM Wiki	350K	101s	16.0
Graphify (GraphRAG)	617K	180s	13.6

Below is the experimental design and detailed numbers that support this conclusion.

A Close Look at Ruflo: What It Means to Layer Multi-Agent Orchestration onto Claude Code

Mon, 04 May 2026 11:30:00 +0900

How to design Claude Code workflows that learn from a team’s insights, not just a single project.

Dohyun Jung - Principal Consultant, ROBOCO

Use Claude Code for a month and the limits become clear. Context disappears along with the session, and breaking work apart and putting it back together is ultimately a human job. The bigger problem surfaces the moment you have multiple projects and multiple developers. The testing strategy learned on project A, the refactoring pattern discovered by developer B, the architectural judgment validated in repository C — none of it flows well between them. ruvnet/ruflo (formerly Claude Flow) targets exactly this gap. As of May 4, 2026, it has 39.4k stars and 4.5k forks on GitHub, with v3.6.27 as the latest release.¹² Ruflo aims to be an orchestration layer on top of Claude Code that adds roughly 100 specialized agents, HNSW vector memory, a plugin system, and Zero-Trust federation. This post lays out what Ruflo actually solves and what it does not, particularly from the standpoint of how to organically reuse the insights that multiple projects and multiple developers acquire.

Serverless Autoresearch: A Record of Making an ML Experiment Pipeline Serverless Through Vibe Coding

Sun, 29 Mar 2026 16:00:00 +0900

The real value of vibe coding is not only in writing code fast. It is in designing a way of operating that lets you repeat expensive experiments more cheaply and more quickly.

Dohyun Jung - Principal Consultant, ROBOCO

Most people first associate vibe coding with web app prototyping or CRUD automation. But ROBOCO’s serverless-autoresearch repository widens that frame considerably. Instead of the “tie up a single H100 for several hours” approach that Andrej Karpathy’s autoresearch presupposes, this project builds a structure that runs parallel experiments in short, explosive bursts on AWS SageMaker Spot.¹

OpenClaw's Five-Layer Gatekeeping: Quality Control in the Age of AI Agents

Fri, 27 Mar 2026 00:00:00 +0900

Thousands of PRs pouring into a 310K-star project, plus parallel commits from eight AI agents. The secret to keeping quality from collapsing amid that chaos is five layers of automated gatekeeping.

Dohyun Jung - Principal Consultant, ROBOCO

In an earlier post I analyzed the productivity principles embedded in OpenClaw’s AGENTS.md. The heart of that piece was “writing the rules down.” But writing rules and enforcing them are entirely different problems.

OpenClaw: Nine Best Practices from the World's Largest Vibe Coding Project

Sat, 14 Mar 2026 10:00:00 +0900

“I used to lead a team. I had a lot of software engineers under me. Even then, I had to accept that they would not write code exactly the way I wanted it written.” – Peter Steinberger¹

Dohyun Jung - Principal Consultant, ROBOCO

OpenClaw is the most-starred software repository on GitHub (310K+ stars) and the definitive case study in large-scale AI-assisted “vibe coding.”² Built by Austrian developer Peter Steinberger using three to eight parallel AI agent instances,³ the project demonstrates that a single developer can orchestrate a 300K+ LOC TypeScript monorepo, more than 20 messaging integrations, and native apps for three platforms — while barely reading most of the code.¹

Everything Claude Code vs Oh My ClaudeCode: A Comparison for Team and Enterprise Adoption

Tue, 27 Jan 2026 09:30:31 +0900

Everything Claude Code (ECC) hands developers a rich set of tools and guidelines. Its approach is to improve the output by making you follow optimal development habits and patterns. Oh My ClaudeCode (OMC), by contrast, automatically orchestrates multiple agents without any complicated setup. It focuses on getting results fast.

Dohyun Jung - Principal Consultant, ROBOCO

I covered each of these in two earlier posts, Everything Claude Code Distilled and Oh My ClaudeCode Distilled. In this post I compare the two tools currently generating the most buzz in the vibe coding community, to help you choose between them.

Oh My Claude Code - The Plugin That Turns Claude Code Into a 'Team'

Wed, 21 Jan 2026 22:03:59 +0900

“Don’t learn Claude Code. Just use OMC.” – oh-my-claudecode README¹

Dohyun Jung - Principal Consultant, ROBOCO

oh-my-claudecode (OMC) is a plugin that layers “multi-agent orchestration” on top of Claude Code.¹ Its goal is to automatically activate the behaviors you need — planning, parallelization, persistent execution, research, design sensibility — using your natural-language request as the cue, so you never have to learn concepts like subagents, skills, and hooks one at a time.

Everything Claude Code: A Hackathon Winner's Recipe for an AI Dev Team

Tue, 20 Jan 2026 09:30:07 +0900

“If you want to bring AI onto the team, you have to design the process before the tools.” – Affaan Mustafa, winner of the Anthropic x Forum Ventures hackathon

Dohyun Jung - Principal Consultant, ROBOCO

Everything Claude Code is a collection of configurations that turns the Claude Code CLI into a virtual development team environment. A hackathon winner spent ten months refining these recipes while building an actual startup product, then gathered them all into one public repository; apply them and you can summon Claude as a “senior engineer + QA + architect.”¹ The repository is open for anyone to inspect on GitHub.