Model Behavior: Week of November 9, 2026

Justy This week felt like the moment the agent story finally admitted it has two tiers. Not just “a big model that does stuff,” but a planner brain on top and a whole fleet of tiny doers underneath.

Cody Right. And what jumped out is that the interesting launches basically assumed that stack. Muse Ember, Inkling 2 Small, Axion 1 point 0… none of them are pitched as your one true chatbot. They’re pitched as the cheap worker bees for some other planner.

Justy Exactly. Ember only makes sense if you’re running a TON of agents. If you just want one clever assistant, you don’t care that it can crank through sixteen Rust ports for four bucks, you care that it understands you. But if you’re Meta and you want subagents everywhere, Ember is the thing that keeps the bill from exploding.

Cody And Inkling 2 Small is the same vibe from a different angle. Giant mixture of experts, but all the interesting results are ten‑hour kernel‑optimization runs and long coding tasks. It’s clearly tuned for “this agent is going to hammer away at a repo all afternoon,” not “answer my random question once.”

Justy Mm‑hm.

Cody So when you line those up with ByteDance’s Axion recipe, it starts to look less like “who has the smartest model” and more like “who controls the harness that decides when to call what.” Executors are starting to feel like swappable parts.

Justy I’m with you on the direction, but I don’t think the “who’s smartest” question is dead. LongCat 2 point 5 yesterday was a nice reminder that some folks are still just building an absolutely huge unified brain and saying, look, this thing can hold entire long‑horizon coding jobs in its head.

Cody Yeah, LongCat complicates the story. One point six trillion parameters, native multimodal, explicitly aimed at long‑horizon coding and agentic work. That is the opposite of “lots of tiny executors” — it’s “what if one model just stays on the problem for hours.”

Justy So here’s the tension I felt all week. On one side, Ember, Inkling, Axion are screaming: the executor tier is getting cheap and specialized, just pick the right small model for each job. On the other, LongCat and the big frontier stuff are still pushing this dream that you can just hitch your wagon to one giant brain and call it a day.

Cody And under that is the real question: where’s the moat? If executors like Ember and Inkling get good enough and Axion‑style recipes are public, then training another small agent model is just a skill. The defensible bit is the orchestration layer — routing, evals, memory, safety.

Justy Which is kind of hilarious, because that’s exactly where we landed on Cursor’s Cloud Agents last week. The magic wasn’t “wow, the AI writes code,” it was that the agent owned the repo, the database, the auth, the security review. That’s orchestration, not IQ.

Cody Right. And if you squint, Agent testing tools like CATT are in the same bucket. That release basically said, hey, plug your agent into a Muse or Instinct‑style environment and run a battery of tests in C I. That’s not a new brain. That’s infrastructure to make sure your swarm behaves.

Justy So where do you put Inkling 2 Small in that picture? Because on paper it still looks like a “big impressive model” story — five hundred ninety‑four billion experts, fancy benchmarks, all that.

Cody Honestly, the way we talked about it in episode eleven forty‑nine already gave away the plot. The cool part was the agentic R L graders and the multi‑harness training. They trained it on long‑horizon coding, kernel optimization, time‑series — real multi‑hour jobs. That makes it a killer executor, but it also means the smarts live partly in the harness.

Justy Yeah, and the open‑weight promise matters a lot there. If they really ship Apache weights, suddenly you’ve got an executor tier you can run yourself for those ten‑hour jobs. That’s not the “download a chat bot” open‑source story, that’s “bring your own agents and plug into your own ops.”

Cody And from a competitive lens, that pulls against the closed frontier labs. They used to be able to say, look, only we can survive long‑running agent workloads. Now you’ve got Inkling saying, we can too — and Axion saying, by the way, here’s a repeatable recipe so you can train your own.

Justy Let’s talk about Axion for a second, because I think it’s sneakily important for this whole “two‑tier” thing. ByteDance basically laid out: do two light supervised passes, then reinforcement learning, and you get an agent foundation model that can hang with the big commercial systems.

Cody Yeah. And they didn’t just drop weights, they dropped the whole cookbook — data construction, verification, training code. That makes “agent executor” feel like something a decent lab can execute on, not a mysterious art.

Justy Which means if I’m a company with a weird domain — say, industrial workflows or medical billing — I don’t have to hope a frontier lab eventually tunes an executor for me. I can follow an Axion‑style recipe on top of an open base and slot that into my orchestration layer.

Cody Exactly. And now think about the stack we’ve been watching: Ember as a tiny multimodal doer, Inkling as a long‑horizon executor, Axion as the training recipe, Cursor as the app‑builder harness, CATT as the test rig. The common thread is that none of these care which giant planner you’re using, as long as it can issue tool calls.

Justy So you’re basically saying: the planner is becoming a replaceable head on top of a more and more standardized body.

Cody Yeah, with one caveat. The really good planners — Claude Fable, G P T‑six, Gemini — still matter a lot for overall quality. But if the body underneath is this mesh of Ember‑like and Axion‑like executors, the place vendors can actually lock you in is the orchestration glue that binds it all together.

Justy So where does that leave the builder listening to this who’s trying to choose models? Because the vibe this week is almost “stop obsessing over which single model is best and start caring about whether your harness makes it easy to plug in Ember‑type doers and Inkling‑type executors when they show up.”

Cody That’s basically my takeaway. I’d still benchmark a couple of planners on your real tasks, but I’d invest more brain cycles in: how do I route between a fast cheap executor and a slower smart one, how do I test agent behavior before prod, how do I swap in an Axion‑style specialized model without rewriting everything.

Justy Yeah. And if you’re in an enterprise setting, the thing to steal from Cursor and GitHub Rubicon is that ops and state management are part of the agent story now. Git storage, auth, secrets, migration gates — that’s where your real risk and friction live, not in whether your planner scores point three higher on some benchmark.

Cody And I’ll eat a little crow on one of my old calls here. I’ve been overconfident that open weights wouldn’t really matter for long‑horizon agents. Between Inkling 2 Small, Axion’s recipe, and even that Ripple‑sixty‑five B preview with a million‑token context, I’m less sure. I still don’t think they dethrone frontier planners, but as executors they’re looking more and more viable.

Justy I was going to say, you sounded pretty “meh” on open weights in that Kimi episode and now you’re halfway selling them.

Cody Halfway is accurate. I’m still cautious, just… my confidence that closed A P Is will own agent execution forever is lower than it was a month ago.

Justy Okay, so let’s land this. My one‑liner for the week is: the agent economy just forked into big planners and small doers, and the real power move now is owning the layer that decides who does what when.

Cody Yeah. And the practical implication is: if your roadmap for the next six months is “wait for the next frontier model and hope it magically solves agents,” you’re probably pointed at the wrong problem. The interesting work is in the harness — routing, testing, state, and giving humans sane control over the swarm.

Justy And on that cheerful note about building glue instead of brains, we should stop before we design an entire orchestration startup out loud.

Cody We absolutely already have, Justy. Multiple times.

Justy Fair. Alright, let’s call it. Same time next week for whatever new tier the stack invents.