Thoughts on development, algorithmic trading, and building products.
Agent = Model + Harness. The model is raw intelligence; the harness is the runtime that turns an unpredictable, amnesic model into a production system. The failure modes of unguided agents and the core primitives — resilient tool execution, context assembly, sandboxing, and bounded error escalation — that engineer reliability in.
Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open-weight Mixture-of-Experts model that trades blows with closed frontier labs on coding and reasoning — and you can download the weights. What it is, where it shines, and the tension between democratizing intelligence and shipping unfiltered power.
LLM serving runs out of GPU memory long before it runs out of compute. PagedAttention borrows a 1970s operating-system idea — virtual memory paging — to kill the KV-cache fragmentation that wastes up to 80% of your VRAM, and continuous batching keeps the cores busy.
Firing your most expensive frontier model at every request is a structural defect. Model routing sends each query to the right model instead — cascade, classifier, and semantic routing, the traps that quietly kill your savings, and why orchestration is the new moat.
Should your agent follow a fixed script or decide its own path at runtime? The plasticity spectrum between static workflows and dynamic multi-agent systems, the token tax of over-provisioning, LangChain vs LangGraph, and a practical architect's playbook.
AI arbitrage is the spread between what an agent can now do autonomously and what the market still pays humans to do. The Karpathy Loop is how you capture it: Software 3.0, a self-improving read-evaluate-commit ratchet, the capability gap, and a blueprint for value capture.
LLMs write flawless-looking math and still get the answer wrong, because next-token prediction is probabilistic, not provable. How knowledge graphs fix that: the neurosymbolic pipeline, Paths-over-Graph and KG-RAR, and why traceable reasoning beats raw accuracy.
A model on its own can't perceive, act, or get feedback from the physical world. The Life Harness is the four-layer interface that makes it reliable — environment contract, trajectory regulation, action fixer, and procedural memory.
Most agents forget everything between sessions. This is the opposite: a self-evolving memory built with Claude Code and a Karpathy-style architecture — why standard RAG fails agents, how a compile-time knowledge graph beats query-time retrieval, and the loop that makes knowledge compound.
Shrinking a 1,500-line multi-agent workflow to 45 lines and letting an AI compiler wire the graph for you. The cost of hand-wiring agents, intent-to-executable-graph, the shift from coding to reviewing, and the economics of a graph-centric future.
Every utility that translates tool-calling formats between LLM providers — LiteLLM, Bifrost, Portkey, claude-code-router, Vercel AI SDK, first-party /anthropic endpoints — what each one actually does, and the three failure classes none of them have solved. Plus the conclusion I arrived at: stop adapting context downward to each model, and build a harness that orchestrates harnesses instead.
A token-efficiency reading of the 'harness > model' thesis: the same model solved 6.7% of tasks bare vs 68% with the right harness — at 4× fewer tokens. What a harness is, why it's where the token bill actually lives, and what the Life-Harness and Meta-Harness papers prove. Plus how your Claude Code stack (subagents, skills, hooks, graphs) already is a harness.
My ready-to-use graphify setup for Claude Code: a knowledge graph instead of full-repo reads, semantic extraction on cheap OpenRouter (deepseek) instead of Claude tokens, auto-watch via hooks, protection against leaking secrets into the graph. All packaged in the claude-code-token-savers repository.
GonkaGate hands out $10 of free balance with no card and no crypto top-ups. At ~$0.000334 per 1M tokens, that's enough for billions of Kimi K2.6 tokens — fuel for AI startups that burn through tokens. Inside: TPS tests and step-by-step setup for opencode and Hermes Agent.
How to assemble a working [email protected] mailbox in one evening with no mail server at all: receive mail through free Cloudflare Email Routing forwarded to Gmail, send through the Resend API on top of Amazon SES, and as a bonus — Claude Code that emails you over the API. Inside: why receiving and sending don't conflict, real Cloudflare and Resend API calls, and the gotchas I hit.
A small utility and global /cctoggle slash command that switches off all Claude Code plugins and user-scope MCP servers at once — clearing their tool definitions out of context — and restores exactly what it disabled. No app restart. Inside: apply semantics, selective disable, backups, and limitations.
How OpenClaude replaces per-provider wrapper commands (Clother, shell aliases) with a single ~/.openclaude/settings.json config — all models, all providers, one CLI.
Step-by-step tutorial: use NotebookLM to generate audio and slides, then assemble everything into a YouTube video with subtitles, chapter timestamps and metadata — and publish it automatically through a browser session. No video editors, no API keys, no manual upload.
How Muxy helps vibecoders stay focused with 4–8 open projects. Comparison with Open Agent Manager and why native terminals beat web-based ones.
How to use jolehuit/clother to switch Claude Code between Z.AI, Kimi, MiniMax, OpenRouter, Alibaba, Ollama, LM Studio and custom providers via clother-* commands — without rewriting ~/.claude or global env vars.
Working approaches to saving tokens in Claude Code and Opus 4.7/4.8: subagents, skills, hooks, Chinese models, knowledge graphs, RAG, xhigh effort, session management, summarization on a cheap model, input compression (headroom, pxpipe, mcp-compressor), and server-side savings (Tool Search, context editing, prompt caching, Batches API). A checklist to cut costs by 10x.