Writing GPU kernels used to be craft work for performance engineers. The efficiency of large language model training and inference often hinges on a handful of critical operators, and who writes them, and how well, directly sets the burn rate of a cluster. The standard move this cycle has been to give AI the full kit: domain knowledge bases, evaluation benchmarks, search workflows, all bundled into specialized kernel agents. JustGoal, a paper submitted to arXiv on October 27, asks the uncomfortable question in reverse: now that foundation models are this strong, how much is all that gear still worth?
The Setup: Take All the Gear Away
The paper's method is deliberately austere. The so-called goal-only baseline hands a general-purpose agent nothing but the optimization objective and constraints — what the operator computes, the input and output shapes, the hardware and performance targets. Everything else is confiscated: no kernel-specific infrastructure, no knowledge bases, no predefined workflows. The agent builds its own evaluation loop: write a kernel, measure, locate the bottleneck, revise, repeat until the target is met.
This drags the control group of every specialized system onto the table. If a bare general agent with a self-built loop can match or beat specialized systems, then the elaborate designs in those papers need to explain themselves.
The Scorecard on B200
The numbers reported in the paper (arXiv:2610.36434) are solid. On NVIDIA B200 GPUs, JustGoal outperforms open-source kernel libraries on 8 of 10 operators, with per-operator geometric mean speedups from 1.14x to 2.34x. Against the fastest publicly released specialized-agent kernel, it wins on 5 of 9 operators with available comparisons. The most telling result is the controlled same-model comparison: both running Fable 5.1 on single shapes, the KDA-0.5 harness with its full stack of specialized machinery reaches only 0.95x of JustGoal's performance. The ablations are equally blunt — adding domain knowledge alone, or predefined workflows alone, shows no consistent benefit.
The Real Payload Is the Methodology
More valuable than the numbers is the paper's conclusion: any specialized machinery must be justified by clear gains over a same-model, same-scope, goal-only baseline, and reevaluated every time the foundation model changes.
This sets a new rule for the entire "agent + domain knowledge" research paradigm. For the past two years, plenty of papers told a story like "we added a domain knowledge base and a search strategy to an LLM, and gained X percent," yet few answered how much of that gain came from the scaffolding rather than from the model simply getting stronger. JustGoal supplies that missing control group. Follow the logic and the moat of a specialized harness comes with an expiry date — every foundation model upgrade forces a fresh re-queue.
The lesson for engineering teams is just as direct: before investing in your own agent workflow, spend a day writing a goal-only baseline and running it. If your specialized system cannot beat the bare agent, the problem is usually not the model; it is the scaffolding.
So — take inventory of your own workflows: which parts are a real moat, and which are just patches for what the base model cannot yet do on its own? The more of the latter, the more you will have to rewrite at the next model upgrade.