Three weeks after my first round of catch-up, two things have given out: my power supply and my patience with the word “agentic”.
First the easy part. I swapped the two Qwen 3.6 27b instances for Qwen3.8 27B, running the Unsloth Dynamic V3.0 GGUFs (UD-Q4_K_XL) with q8_0 KV cache quantisation on both Ollama instances, at 196k context. Unsloth uses a dynamic quantisation approach, worth a read.
Then I ran both cards flat out for extended periods. It turns out a be quiet! Straight Power 11 Platinum 850W is perfectly for two PowerColor R9700s in bursts, and much less happy when they are pinned for a longer stretch alongside a Ryzen 9 5900X. No bang, no smoke — just a distinct tic and a black screen of death. Ordered a Straight Power 12 Platinum 1200W at Alternate, mostly for the two native 12V-2×6 connectors. My earlier worries about hot feet and noise-cancelling headphones stand.
Which brings me to the actual point of this post.
The setup is the same as before: the two local Qwens do the typing, Claude does the thinking and the quality control. And Claude, the supposedly agentic bit at the top of the stack, keeps walking into walls that only a human with subject-matter knowledge can see coming. Not syntax errors. Not hallucinated libraries. Architectural road blocks, missing functionality (there is a user-management screen, please create the navigation to get to it from the management landing page) — the unreachable new page delivered with complete confidence.
My favourite example, because it is so beautifully wrong: at some point Claude decided we needed to prove ownership of the tenant’s domain, and proposed doing it the way every getting-started guide on the internet does it: drop a TXT record in DNS and wait for propagation.
Except that we had, together, in the same session, already set up an Entra OIDC integration. Which is to say: an app registration in the very tenant that owns that domain, issuing tokens for actual, named users authenticating against their IdP. The trust relationship was not only already established, it was established considerably more convincingly than a TXT record ever could. A DNS record proves that somebody, once, had write access to a zone file. An IdP federation with the tenant proves you are talking to the organisation and to its people, continuously.
Claude did not actively know that. Or rather, it knew both facts and never joined them up. Its algorithm correctly identified the need to “verify a domain”; unfortunately, it failed to understand it had already satisfied the requirement. Someone coined the phrase “imitating intelligence” and that is the best definition I have heard so far (from a podcast, forgot which one) for the seemingly incredible amount of knowledge the LLM’s have, but they can not be relied upon to connect all the dots.
And this is where I get slightly cranky about the Agentic claim. We have gone from Vibe-coding to Agentic Engineering in about a year, and the second term is meant to signal that the machine now has agency: it plans, it decides, it executes. What I actually have on my desk is a very fast, very well-read team of juniors who produces more code before lunch than I do in a month, and who will confidently take a wrong architectural turn, forget about its audience, or implement completely useless functionality.
Learning points
- Agentic engineering does not remove the need for deep understanding of the tech being used. I made my orchestration agent quite chatty, and the questions it asks usually have deep repercussions for the usability and maintainability of the system.
- Model quality is not the bottleneck any more. Qwen3.8 writes fine code. The bottleneck is that nothing in the stack holds the architectural intent between steps.
- Sustained load is a different beast from benchmark load. Size your PSU for the hour, not for the demo. 850W was plenty right up until the tic.
- The failures that cost you money are not the ones that throw errors. Broken code announces itself; a redundant-but-working design decision quietly lives on for years.
- Write documentation, and re-write it. I have installed Docmost (wiki) and Plane (backlog/project management) and have a specific agent skill for (re-)writing all documentation. I spent about 40 hours with Claude creating the architectural and functional documentation before any code was written. This sounds limited and it is, because I already did the work for this project and still it took 40 hours to write it down again.
Next steps
The 1200W supply goes in today, together with a second NVMe so the models stop fighting my boot drive for space (which, delightfully, means pulling a GPU to reach the M.2 slot — whoever designs these layouts owes me a beer).
Btw, if you need someone to lead your developers — be it human or AI — in delivering custom applications: I am available for a new assignment.