(that's a Judas Priest reference, btw)
Cam. Compiled from ten days of build logs, 2026-04-30 → 2026-05-09.
I set out to get a small autonomous monitor-and-react agent off the ground — on a DGX Spark / GB10 / ARM64 box, serving open-weight models locally, with the goal of eventually hosting dozens of agentic workflows authored by non-technical family members. I'd POC'd the usual alternatives (n8n + Ollama, LangChain, LlamaIndex) but the shiny Spark/Blackwell hardware plus the hype around OpenClaw/NemoClaw pulled me in.
The first writeup was the morning-after read on standing the stack up. What follows is ten days: the stack came up, the model lied to me, the framework turned out to be the thing lying, I swapped the whole UX layer, and by the end there was a real product monitoring real feeds. This is the honest version — what's good, what's bad, and what's ugly, updated as the picture kept changing under me.
- OpenClaw + NemoClaw — the framework read — the architecture is genuinely well-designed, the maintainers actually engage, and the defaults are tuned for frontier models in ways that quietly wreck local inference. Where I started.
- The model question — Hermes-3 hallucinates tool calls it never made. Qwen3.6 fixed most of it, Qwen2.5:7b turned out faster and honest, and the MoE-is-faster hypothesis was wrong by an order of magnitude. Plus why DeepSeek-V4 doesn't fit the box.
- The pivot — the framework was the source of the worst failure mode, not the model. A LangGraph spike proved it in one evening, and the multi-tenant UX problem got solved by not writing a frontend at all.
- Building the product — retrieval quality is a three-layer problem, a cross-encoder reranker bought a 30× speedup, and the CVE warmup loop transferred almost verbatim into the real email-triage pipeline.
- The cluster grows up — a two-instance prod/dev split, a fifth host, and a runtime bench that came out 3–6× off from what I expected.
- Did anyone fix it? — the upstream scorecard — two months later, which bugs actually moved. NemoClaw fixes things, OpenClaw labels things and stalls, and Dify's community fixes things while the merge waits on code owners.
Aside — Atlas: going deeper into the machine. A separate adventure off the main arc: opening up an open-source Rust inference engine and running a 122B mixture-of-experts model across two GB10 boxes with expert parallelism. The one stretch where I got under the abstractions — MoE kernel selection, quantization boundaries, speculative decoding, CUDA graphs, and the bandwidth wall — in the language it's actually implemented in.
OpenClaw is closer to "production-ready for technical users" than I expected, and further from "production-ready for non-technical family members" than I'd hoped. Every bug I filed had a clean structural fix on the table. Nothing in flight closed the gap inside a week.
The two decisions that mattered: move production off Hermes-3 (it fabricates tool calls with total confidence), and stop trying to hand-author skills as bare framework config. The second one is what pushed me from bare OpenClaw → a LangGraph spike → hosting the whole multi-tenant surface on Dify, wrapping my own retrieval skills behind an OpenAPI bridge. That let me keep the parts of OpenClaw/NemoClaw that were good (the observability discipline, the testing muscle) without betting the family-facing product on the parts that weren't ready.
By day ten there's a real thing running. A retrieval agent with a cross-encoder reranker, a CVE-advisory monitor that matches new entries against a tracked-tech list, and an email-triage pipeline that classifies and drafts replies — the actual monitor-and-react shape I was after, on local inference, on hardware I own.
The upstream verdict, two months on, is the tiebreaker. I filed across all three projects and let the results come in (the scorecard has the receipts). NemoClaw fixed the load-bearing sandbox blocker and merged a three-in-one reliability PR — that's a project you can build on. OpenClaw's bot triages beautifully and then stalls: P1 security bugs sit stale, a validated filing got auto-closed, and my complaint about a bad bot-closure was closed and locked with no human reply. Dify's one bug drew a community-written fix PR that's still waiting on code-owner review. Responsiveness, not architecture, is what set the timeline I could actually plan around — and it's most of why the product routed off OpenClaw's own roadmap.
What I'd tell someone evaluating this stack today: file your bugs — but know that "the maintainers will engage" is true for NemoClaw and only half-true for OpenClaw, where a beautifully-labeled P1 can sit untouched for months. Don't run local inference on Hermes-3 with a tool surface. Expect to rewrite the default AGENTS.md. And treat "the agent said it did the work" as a claim to verify at the wire, not a fact — on a small local model it is the common failure case, not the rare one.
Filed upstream across the run (OpenClaw, NemoClaw, and langgenius/dify): 15+ issues and evidence-comments. NemoClaw shipped several as merged fixes; most OpenClaw filings remain open or bot-closed; Dify's has a community fix PR in review. The scorecard tracks each one's current state.