Some thoughts on the Agents A1 model.
I’ve been testing it against Qwen 3.6 and Qwen 3.5 122B on coding-agent workflows, mostly to get a feel for where it actually lands in practice.
My current read:
The interesting part is where it succeeds and where it falls apart.
Unlike Qwen 3.6, Agents A1 passed my evaluation prompt in Go. It did not pass in Zig.
In fact, it does not seem to have the expertise to build even a basic API in Zig. I tried.
The Zig attempt produced a recursive rerun effect, where the agent kept looping back over the same sections of code again and again, without really advancing the implementation. That is usually a bad sign. It means the model has lost the thread, but still has enough confidence to keep moving the furniture around.
It also failed on both the Zig and Next.js sides of another test in a way that was pronounced enough that I would not currently trust it for either stack.
The Next.js failure was a different kind of problem.
It appeared to pattern-match against both older and newer documentation. The agent did seem aware that there was a version divide, and even mentioned that distinction in its reasoning, but it could not reliably navigate through the problem without explicit prompting.
That is good to know.
A second, better-bounded test
So I adjusted the experiment.
Same general task, but with a more plausible success state: Go/Vite instead of Zig/Next.js.
The prompt was simple: build a basic ecommerce website and catalog, with no payment processing.
First round: broken.
Not surprising. Most models break something on the first pass.
The second step in both versions of the experiment was to use Codex or Claude for code QA, evaluation, and re-prompting.
Codex handed in its notes.
Agents A1 made the repairs in one shot, without asking for clarification.
That part impressed me.
My current takeaway:
Both Agents A1 and Qwen 3.6 are heavier than expected. They are both long-context MoE models with real reasoning ability, and they both feel more capable than I expected going in.
But based on this evaluation, Agents A1 is the faster and better model on my Nvidia DGX Spark.