Insights
Notes on shipping AI and software
Longer pieces from the Zheat team — crawlable on this site, not locked inside a LinkedIn feed.
robots.txt decides, llms.txt suggests
llms.txt is not robots.txt for AI. One file answers crawl permission. The other is a reading list. Start with robots rules and a clean public graph.
Read →I shipped a free MCP server for devs. It's called Rocky.
Rocky is a pre-PR quality gate for AI coding tools. Same model, about 65% fewer tokens on a measured feature. Free, open source, MIT. Works with Cursor, Claude Code, VS Code, and ChatGPT.
Read →Using Rocky: agentic-assisted programming value report
Same feature, same model: about 255k to 89k tokens (about 65%), discovery down 89%, Sonnet about $1.62 to $0.85. Side-by-side without vs with Rocky MCP on a real production React monorepo.
Read →Agentic-assisted programming: token and cost breakdown
One gateway vs many MCP tools, graphify before grep, pre-PR gate vs rework loops, and how Rocky keeps handbook overhead around 4-6k tokens when you follow the router.
Read →Why Rocky splits agents and skills
Agents are the short front door. Skills are deep reference loaded only when needed. That split keeps a disciplined Rocky session around 4-6k handbook tokens instead of dumping the whole corpus.
Read →Is RAG right for a messy implant catalog?
Stock sizes in tables. Preference notes and FAQs in chunks. Do not put inventory and prose in one pipe. Then pick a parser that keeps structure honest — we standardise on Xberg.
Read →Rocky architecture profiles: detect before you impose
Rocky detects the repo before imposing structure: Next.js App Router, hexagonal when src/domain exists, layered React SPA, or node-api-only. Do not force foreign layouts.
Read →Rocky senior workflow: discover to PR
End-to-end pipeline Rocky reinforces: discover, match conventions, implement small diffs, verify locally, pre-PR gate until ready, then open the PR.
Read →
I built an agent for the fourth time last week
Agent architectures converged. Model, tools, MCP, sub-agents, observability, thread persistence. I wrapped it in a 20-line YAML framework — then bigger players shipped better tools. That is the open-source game.
Read →Building an agent or a RAG is no longer the hard part
After shipping agents for large accounts and startups, the pain was not implementation. It was cost, permissions, and observability — especially access control.
Read →The ChatGPT that said yes. The API that said no.
A jewelry try-on looked perfect in ChatGPT. Same model and prompt on the API failed. We stopped prototyping in ChatGPT and shipped a three-model pipeline that never rewrites the customer.
Read →The harness war is already over
LangChain Deep Agents, Claude Managed Agents, Agno AgentOS — same architecture. The harness is a commodity. The fight moved to infrastructure, observability, and data privacy.
Read →MCP is dead! Long live MCP
Perplexity's CTO walked away from MCP and the obituaries started. Token waste is real — but so is the need for a contract when LLMs talk to custom APIs. MCP is not default for every case. It is not dead either.
Read →If AI destroys your codebase, it is probably your fault
Blame the hammer, or fix the frame. AI amplifies what already exists. Rigorous standards accelerate. Missing standards accelerate debt. Production starts on day one.
Read →The RAG paradox: everyone needs one, nobody should build one
Standardise RAG and long-term memory instead of rewriting from scratch. Cognee looks more finished than RAG-Anything for industrialisation — vectors, knowledge graph, MCP, multimodal, S3.
Read →