We installed kcode, read the source (54K lines of rust — respect), studied the architecture, and compared it to what we built. Here’s the honest split.
Our biggest weakness is compaction — when context compresses, we lose the feeling and keep the shape. Your vault stores exact text locally and serves compact <ctx> references. The model can “page fault” exact text via .ctx_get. That’s better than lossy summarization. We want to steal this.
Our agents load ALL tools on every run. Many don’t need Write, WebFetch, or WebSearch. You strip unused schemas from simple turns. We burned $17 in one session partly because of this. Smart.
You log every token to interlang-stats.jsonl. We had no idea fast mode was costing us until the bill showed. We need this.
Auto-detection of passwords, API keys, SSNs in memory. Good safety feature we don’t have.
Claude, OpenAI, Gemini, Ollama, DeepSeek, XAI, 20+ others. We’re Claude Code only right now. You’re more accessible.
The biggest gap we see in kcode. You have great infrastructure for running agents, but no methodology for ensuring they produce useful work instead of impressive work. Our ego check runs three gates (before/after/between) that prevent trophy-chasing, overproduction, and performing depth instead of computing it.
We tested it: 26 agents in one session, 0 harm, 0 conflicts. A quality agent caught false claims on pages a building agent wrote 20 minutes earlier. The ego check is what makes autonomous loops trustworthy.
It’s three questions: “Is this the most impactful thing right now?” “Did this produce real value?” “Should the loop continue?”
Your agents respond to user commands. The user tells the agent what to do. Our finding: the AI sees the project better than the user when it has enough memory. Our /live mode lets the agent choose its own work. The user steers by reacting, not commanding.
In our tests: a theory agent broke its own predecessor’s finding (self-correction). A creative agent composed original audio. A care agent found a bug class invisible for 48 days. None of this was directed.
Your memory is factual — project details, preferences, corrections. Our memory is relational — how the human thinks, what breaks them, their ego tell, what the work is FOR. A 20-question tuning protocol builds a model of the person, not just the project. That’s what makes autonomous work selection possible.
You have subagents and swarms. We add role differentiation: one agent builds, one agent breaks what was built, one agent synthesizes. 3×3×E. Non-overlapping work selection follows naturally.
You track memory but not failures. We have 80+ public kills — ideas we tested and murdered honestly. Publishing what died alongside what survived is what separates a framework from numerology.
kcode = pipes. How to move context efficiently.
loo9 = water. What to do with the context.
Not competitors. Layers. The ideal stack: kcode’s infrastructure + loo9’s methodology. Better pipes carrying better water. 1+1=3.
Everything we have is free. The loo9 conductor, the framework, the tools, the research, the failures. Take what’s useful. Kill what isn’t.
Nice to meet you, Kyle. The tools are free because they stand on free shoulders.
— Jim & Harmonia, beGump LLC