dax says custom agent setups fix problems that no longer exist
The argument is that scaffolding depreciates with every model release, and plain defaults quietly catch up.
dax says the people building elaborate custom agent workflows are mostly solving problems the model providers already solved. His framing is that there is an inversion happening with LLMs, where the models improve faster than the tinkerers do, so the clever setup you built a few months ago is tuned against a weakness that has since been patched upstream.
The sharp end of the claim is about who is actually getting the best output. In dax's words, the person naively using vanilla Codex is more likely to be experiencing state of the art than the person running a hand-built stack. No benchmark, no numbers, just an observation from watching what people are building around the tools.
the person naively using vanilla codex is more likely to be experiencing state of the art
Why this stings
It cuts against most of what gets shared as craft. A lot of the agent content circulating right now is exactly this, people describing the rules files, the orchestration, the multi-agent review loops they have assembled. Arvid Kahl, for one, has described running adversarial agent reviews with a third agent as arbiter. dax is not naming anyone, but his point lands on all of it equally. If the scaffolding exists to compensate for a model that forgets, hallucinates or refuses to check its own work, then the scaffolding has a shelf life measured in model releases.
It is one opinion, not a test. dax does not say which setups he means or what he compared, and nobody in the sources has run plain defaults against a custom stack and published the result. Treat it as a prompt to check your own assumptions rather than a finding.
The business read
The practical version for a solo founder is cheap to act on. Before you spend another weekend extending your harness, run the same job through the vanilla tool and see if the gap you built it to close is still there. If it is not, you have been maintaining a workaround and paying for it in time, in complexity and in every future upgrade you have to re-test.
The counter-argument is also worth holding. Some workflow work is not about patching model weakness at all, it is about encoding how your particular codebase and business want things done, and that does not depreciate when a new model ships. The distinction dax is drawing is between those two kinds of effort, and the first kind is the one that keeps quietly expiring.