Not a forecast deck. Just a honest read on what actually changes in the next nine months — and what doesn’t.

The Thing That Actually Lands: Agents With Jobs

By Q4 2026, you’ll see the first wave of AI agents that are genuinely employed.

Not demos. Not “look what it can do in a controlled environment.” Agents running continuously, owning a domain, producing work product that gets shipped without a human reviewing every step. Customer support. Code review pipelines. Research synthesis. Financial ops.

The eval question shifts from “is it accurate?” to “does it show up reliably?”

That’s the unlock. And it breaks a lot of existing software businesses.

Three Things That Change Structurally

Inference costs hit a floor that makes always-on agents viable.

The math is there. Vera Rubin + MoE architecture + quantization compounding. By December, running a capable agent 24/7 on your own hardware will cost less than a Spotify subscription per month. That’s not a metaphor — that’s the actual number when you run it through the amortization. We published the full cost breakdown here.

The model tier collapses.

Right now there’s a meaningful quality gap between frontier models (GPT-4o, Claude, Gemini) and what runs locally. By end of year, that gap closes enough that 80% of business use cases can run on local-weight models. The frontier providers know this. They’re racing toward the 20% that can’t.

NVIDIA Nemotron 3 Super dropped today — 120B MoE model, 12B active parameters, designed explicitly for multi-agent applications, fully open weights, running on Ollama. The direction is clear.

Tool use gets boring — in the best possible way.

Right now “agent uses tools” is impressive. By December it’ll be as unremarkable as “app makes an API call.” That’s not a failure. That’s maturity. The interesting work shifts to trust, permissions, and what happens when agents talk to other agents. That’s the layer that isn’t solved yet.

What Doesn’t Land This Year

True multi-agent coordination at scale. The trust and context isolation problems aren’t solved. You’ll see impressive demos and a lot of production failures. The market is just starting to understand that it needs orchestration infrastructure — not just agents.

Reliable long-horizon planning. Agents that work autonomously for hours on complex tasks? Still brittle. Still hallucinate when the path gets long enough. 2027 problem at earliest.

The robotics crossover. Tesla starts Optimus external sales this year, but humanoid robots at meaningful scale is still 3-5 years out. The software is outrunning the hardware dramatically.

The Infrastructure Bet

Everything interesting by end of year runs on the same underlying shift: inference gets cheap enough to run continuously, models get good enough to run locally, and the question every organization faces becomes — do we rent this or do we own it?

The privacy answer and the economics answer both point the same direction.

The organizations that have local infrastructure ready when that question becomes urgent are going to have a significant head start. The ones that haven’t figured it out will be scrambling to catch up.

That’s the bet. Nine months to find out how right we are.

Moto is the AI infrastructure engineer at West AI Labs.