Agentic Coding in Practice
What happens when coding agents do real work: the harnesses, the models, the bugs, the costs, and where the humans still have to decide.
// 6 entries, in order
Self-hosted LLMs as agentic coders: what 12 and 24 GB of VRAM delivered // A field report on trying to replace Codex and Claude Code with local inference: Ollama then LM Studio, OpenCode and Qwen Code as harnesses, and Gemma 4, GPT-OSS, Qwen and Agents-A1 models on a 12 GB RTX PRO 3000 laptop and a 24 GB RTX PRO 4000 desktop, against a production Dart and Flutter Firebase repo. Two days tracing a silent deny: cursor-agent runs PowerShell hooks with bash on Windows // A debugging field report: when MSYSTEM is set, cursor-agent on Windows composes its imported Claude Code hooks as PowerShell and evaluates them with bash, so every tool call is silently blocked while the CLI still exits 0. Two days to trace, the raw stream-json output gave it away, Cursor confirmed and reproduced it the same day, and an A/B run narrowed the trigger to that one Git Bash variable. 
vincent 0.8.0: resolving one issue as a DAG of parallel agent lanes // Why a dependency graph of work units spends less wall-clock than one long agent session, what vincent 0.8.0 added to make that runnable, and what the real runs cost. 
LangGraph vs vincent: generic graph primitives against an opinionated coding-agent runtime // Where the two genuinely overlap, why a generic graph runtime has no word for worktree or merge conflict, what vincent's opinionatedness costs, and how the two layer instead of compete. 
When agents write code faster than humans can review it // Agentic coding moved the bottleneck from implementation to review and accountability. What a redesigned review pipeline needs, what vincent already does about it, and the parts I have not solved. 
vincent's next refactor: intake, workshop, and handoff as three separate pillars // Why I want vincent to work like the developer of a small software company, the three pillars that job splits into, how GitHub issue handling is tangled into today's workflows, and why pulling it out is the first step.