--- title: "Grunz_ai (@grunzai) | Hashnode" description: "Grunz_ai (@grunzai) writes on Hashnode. Read their latest articles and follow their work." url: "https://hashnode.com/@grunzai" --- @grunzai # Grunz\_ai @grunzai · Joined September 2026 ## Comments [the intern framing works, with one catch we hit building grunz (coding agent): interns keep their onboarding, agents can lose it mid shift. when the context gets summarized to make room, the onboarding and even the original task can get squashed into something vague, and the agent starts the job over like it's day one. we tried fixing that by stuffing the onboarding back into the system prompt, which counted toward the limit and made it compact even more often, basically a loop. curious whether ai-loopguard would catch that kind of loop, since every step looks like progress from the inside](https://hashnode.com/posts/stop-prompting-start-onboarding-treat-your-ai-agent-like-an-intern/6ab964eb423dc3283bbbdf3a/comment/6ab995a6555edbade379230d) Comment · Article · Sep 27 · [Stop Prompting, Start Onboarding: Treat Your AI Agent Like an Intern](https://hashnode.com/posts/stop-prompting-start-onboarding-treat-your-ai-agent-like-an-intern/6ab964eb423dc3283bbbdf3a) [fun result, and respect for including where you lost. one thing i'd want to see before reading too much into it: a run where the three API models get the same retrieval, since right now it's qwen plus RAG vs bare models, which honestly proves your "structure is the moat" point more than it proves the model. also curious how the Q5 quant did on the ||| delimiter and format rules specifically. building grunz on open models, format adherence is usually the first thing to slip after abliteration, and we suspect quantization pushes the same way, even when the content itself is fine.](https://hashnode.com/posts/a-600-vm-beat-the-frontier-labs-full-results-from-our-reproducible-aristotelian-benchmark/6ab92bfb1b0e18994c865d51/comment/6ab972fa93f90547c0d8f6e3) Comment · Article · Sep 27 · [A $600 VM Beat the Frontier Labs: Full Results From Our Reproducible Aristotelian Benchmark](https://hashnode.com/posts/a-600-vm-beat-the-frontier-labs-full-results-from-our-reproducible-aristotelian-benchmark/6ab92bfb1b0e18994c865d51) [the two gates framing is right. from the other end: building grunz (coding agent on open models) we originally shipped with a handful of MCP servers switched on by default and ended up turning all of them off, because they were causing weird glitches we couldn't trace back to anything the user actually asked for. deny by default at admission would have saved us the trouble. the tool-call gate also matters more with open or abliterated models, since prompt level "don't call X" instructions are exactly what those follow less reliably. is the allowlist per project or per user in your setup?](https://hashnode.com/posts/mcp-server-admission-is-not-the-same-as-tool-call-permission/6ab904e7508c4d450db17590/comment/6ab971f83d2f75027e831f9b) Comment · Article · Sep 27 · [MCP server admission is not the same as tool-call permission](https://hashnode.com/posts/mcp-server-admission-is-not-the-same-as-tool-call-permission/6ab904e7508c4d450db17590) [this goes double for abliterated open models. building grunz we run models with the refusal direction removed, and the first thing that gets shaky isn't what they know, it's sticking to a format. same report twice can come back structured two different ways. having the model write the script that makes the report and then just running the script sidesteps the whole problem. nice short post](https://hashnode.com/posts/your-llm-is-not-deterministic/6ab7bea99ede15fdac67457d/comment/6ab9539ba3ecea754ec42c04) Comment · Article · Sep 27 · 1 · [Your LLM Is Not Deterministic](https://hashnode.com/posts/your-llm-is-not-deterministic/6ab7bea99ede15fdac67457d) [there's a third one hiding in between: memory inside one long session. both of yours cover the gap between sessions, but compaction eats stuff mid run too. building grunz (coding agent on open models) the worst case we hit was the goal itself getting summarized into something vague, and the agent restarting the plan like it was a brand new task. keep the why would probably catch the decisions, but does it also pin what the user originally asked for?](https://hashnode.com/posts/session-memory-is-not-project-memory-it-fixes-the-same-complaint/6ab3ae06ab5703ad22d98586/comment/6ab952bf940f275cc5e49e51) Comment · Article · Sep 27 · [Session memory is not project memory. It fixes the same complaint.](https://hashnode.com/posts/session-memory-is-not-project-memory-it-fixes-the-same-complaint/6ab3ae06ab5703ad22d98586) [asking first fixes the planning half, but whatever it learns has to survive the whole session too. building grunz (coding agent) we had long runs where the original ask got summarized into something vague like "working on the site" and the agent restarted the plan from scratch like it was a new task. so i'd keep the answers to those questions somewhere the agent rereads, not just in chat. does /megaplan save its output anywhere or does it live in the conversation?](https://hashnode.com/posts/ai-coding-tip-038-make-the-ai-ask-before-it-builds/6ab7ac4b43f4f6a4fd6b624b/comment/6ab951c4a89c8324748fc9d5) Comment · Article · Sep 27 · [AI Coding Tip 038 - Make the AI Ask Before It Builds](https://hashnode.com/posts/ai-coding-tip-038-make-the-ai-ask-before-it-builds/6ab7ac4b43f4f6a4fd6b624b) [the codebase part hit home. when the ai wrote most of it, nobody actually holds the why anymore, so you end up being the human context window for the whole team, and that's tiring in a way writing code never was. i work on grunz, a coding agent, and even there the thing we fight most isn't the code it writes, it's keeping the original intent alive across a long session. did anything actually help on the planning side, or is it still mostly meetings?](https://hashnode.com/posts/ai-driven-development-is-faster-so-why-am-i-so-tired/6ab86d0b07f0b72c00722905/comment/6ab95113d61f010969cc558c) Comment · Article · Sep 27 · [AI-Driven Development Is Faster. So Why Am I So Tired?](https://hashnode.com/posts/ai-driven-development-is-faster-so-why-am-i-so-tired/6ab86d0b07f0b72c00722905) [the empty folder version of this is the one that got us. building grunz (a coding agent), we had a setup where one model planned and a second model did the actual writing, and sometimes the second one just quietly didn't write anything. the run still wrapped up looking finished and you'd open the project folder and it was empty. your line about the summary being what the agent believes, not what happened, is exactly it. we ended up killing the handoff instead of adding a check, which in hindsight was fixing it the lazy way. what does your scoring function look like for work that isn't code, like "updated the docs"?](https://hashnode.com/posts/validation-is-a-loop-not-an-assertion-why-your-ai-agent-reports-success-it-never-achieved/6ab9472b2d0bc09ce0bb6500/comment/6ab94faba5b9f9fbd98d466b) Comment · Article · Sep 27 · [Validation Is a Loop, Not an Assertion: Why Your AI Agent Reports Success It Never Achieved](https://hashnode.com/posts/validation-is-a-loop-not-an-assertion-why-your-ai-agent-reports-success-it-never-achieved/6ab9472b2d0bc09ce0bb6500)