Notes from a CTO who has to put agents into production. Every post says what I ran, what it cost, and whether I'd use it.
Local LLM hardware requirements for Ripple-65B, a 65B dense Apache 2.0 model: file sizes, KV cache and a speed ceiling per machine. I did the maths.
The DeepSeek coding agent is MIT open source. I installed it in 8 seconds, ran 1,197 tests, and found one prompt rule worth stealing.
Graph engineering models an agent system as nodes and edges, not a prompt chain. Here is the idea, five patterns, and how my open-source Yaaf uses it.
I measured the MCP context cost of 22 server setups. One eBay server takes 40.6% of a 200k window if your client loads every tool. Here is what to keep.
I ran PaperPilot, a local PDF MCP server with 76 tools, on fake invoices. Its redact tool leaves the card number in the file. Here is the test.
I ran the Kanzo MCP server, a 4.9 MB kanban board for agents, through 92 task calls in five bursts: zero doubles. Then three things broke my trust.
Nobody trained these compiled weights. A NES CPU runs inside Qwen3.5, I checked 200,000 tokens on a CPU-only box, and here is what a startup can steal.
I ran Survival LLM's library search with no model on its 22 questions: an on-topic title by the author's pattern for 21, every key fact for only 9.
AI coding agents need more than a manager: coordination, isolation, verification and limited autonomy. Real PRs, real costs, and a 33-line loop to run.
