I moved my coding agents off hosted APIs and onto two machines I own, a MacBook Pro M5 Pro as the client and a gaming PC with an RX 7800 XT as the model server. Model capability was rarely the blocker. The blockers were unified memory limits, inference engines that mangle tool call formats, a model stuck in a loop, and a laptop that kept freezing for reasons I still have not pinned down.
How I built an agentic writing pipeline on Mastra: researcher, writer, editor, human approval, and the token-limiter bug that made revisions invent details.
Colibrì, a ~1,300-line pure C inference engine with zero dependencies, runs the 744B-parameter GLM 5.2 MoE model on a laptop with 25 GB of RAM and no GPU by streaming int4-quantized experts from NVMe. Cold-cache throughput is 0.05 to 0.1 tokens per second, so speed is the honest limiter, but it is an existence proof that frontier-scale local inference works on commodity hardware.
Coding agents get better every month, and the way to stay in control is to give them clear instructions and context rather than fight the shift. This covers agent instruction file conventions worth adding to a repo (AGENTS.md, SKILL.md, CLAUDE.md, Cursor skills) and two open-source tools, Ponytail and Caveman, that cut token spend by trimming what agents write and compressing what you send.
Model choice is an engineering decision that shows up in both output quality and your bill. This walks through models.dev and artificialanalysis.ai as the two resources for shortlisting on specs/price and then checking capability, cost per task, and latency, with worked comparisons across GPT-5.6 Sol, Claude Fable 5, Gemini, LongCat-2.0, and Seedance 2.0.
OpenAI's GPT-5.6 announcement is dense with benchmark numbers and efficiency claims. The company positions Sol as the flagship model, with Terra and Luna filling in for balanced and cost-efficient tiers. The through line is not just higher raw scores. It is capability per token and time to result across agentic workflows, coding, browsing, science, and cyber tasks, with repeated emphasis on lower cost and latency for comparable outcomes.
Google quietly published OKF, a plain markdown-plus-frontmatter format for agent knowledge, built on a pattern Karpathy sketched in a gist months earlier. Working through both taught me to stop thinking of an LLM as one brain and start thinking of it as two.
Claude Fable 5 was pulled from every user globally for two and a half weeks after a government export control order, then came back with a new safety classifier. Here is what happened and what it tells me about how "free" a model actually is.
I followed the Mastra tutorial to build and publish a simple MCP server, then connected it to a self-built agent and to Claude. Here is what the protocol actually does under the hood and what surprised me.