Why we're betting AI agents go local before they go cloud
· Avery NXR
Every other agent platform is built cloud-first. We built Avery NXR local-first. That decision was deliberate, and we want to be transparent about the bet underlying it — and the way we think this plays out.
The bet
Operational AI agents — the kind that handle real recurring work inside companies — are going to run locally faster than the broader AI category recognizes.
Not "eventually." Not "in five years." Within the next 18-24 months, local execution will be the production-default for operational agents, with cloud LLM access as the opt-in escalation for the rare frontier case.
This is the opposite of the trajectory most of the AI agent industry is on, which assumes cloud-first and treats local as a niche enterprise configuration. We think the industry is wrong.
Why we think we're right
Three things are happening at the same time:
The models crossed the threshold. Qwen 2.5 Coder 7B benchmarks above GPT-4 at launch on coding tasks. DeepSeek-R1-Distill 8B has strong reasoning. Llama 3.x, Mistral small models, the various distillations from frontier labs — all in the same territory. For operational AI work (narrow, repetitive, well-defined), the small specialized models match or exceed cloud frontier models. This wasn't true 18 months ago. It is now.
The hardware is already deployed. Apple Silicon Macs from M2 onward run 7B models comfortably. M3/M4 run 13-14B models. Mid-range PC laptops with 32GB RAM handle the same workload. Most businesses already own this hardware for their employees. The deployment target is in place.
The tooling matured past the usability threshold. Ollama is install-and-forget. Vector databases are embeddable. Agent frameworks support local model deployment natively. The infrastructure that used to require an ML team to deploy now ships in desktop applications.
When all three thresholds cross simultaneously, the architecture that's been theoretically interesting for years becomes practically default.
Why the industry hasn't caught up
The industry is slow to update for a few reasons that aren't malicious — they're just slow:
The viral demos all happen on cloud frontier models. Local AI doesn't make for as good demos. Demos shape the visible understanding of what's possible.
The sales motions are built around cloud SaaS. Field sales teams know how to sell hosted agent platforms. They don't yet know how to sell local-first.
The largest agent vendors built their architecture before the local model thresholds crossed. They've architected around cloud LLM dependencies and can't easily reverse that decision without rebuilding the product.
The buyer mental model is "AI = cloud" because that's the only AI they've ever bought. Updating the mental model takes time.
None of these slow the underlying technical reality. They just slow when the market acknowledges it.
What we'd be wrong about
Two scenarios where we lose:
Scenario 1: Frontier models keep pulling away from small models faster than small models can specialize. If the gap between GPT-7 and Qwen 4 widens instead of narrows, the cloud-frontier-default keeps making sense even for operational work. We don't think this is the trajectory — specialization beats scaling for narrow workloads — but it's possible.
Scenario 2: Cloud-LLM costs drop fast enough that price stops being a meaningful axis. If the per-token cost of frontier models drops 10x in the next two years, the economic argument for local fades. Privacy, latency, and ownership arguments would remain, but cost would stop driving the conversation.
Both scenarios are possible. We think the first is unlikely because of the narrow-task specialization math. We think the second is possible but takes longer than two years to play out.
If we're wrong on both, we lose the bet. The team that built local-first ends up retrofitting cloud-first the way the cloud-first teams are slow to retrofit local. That's the risk we took.
Why we took it anyway
The bet isn't symmetrical.
If we're wrong, we built a local-first product in a cloud-first market. We can add cloud capabilities (Consult Mode already does this for the frontier-escalation case). We're a year or two behind where we'd be if we'd built cloud-first.
If we're right, we built a product that lands ahead of the architectural shift. The teams that have been waiting for "AI tools that don't compound our cost / privacy / sovereignty problems" find us when they go looking. The category moves toward us.
The asymmetry favors taking the bet.
What it means for what we build
The bet shapes every architectural decision:
The model runs on the user's machine. Not on our cloud. We don't have a cloud.
The data stays on the user's machine. The 7 agent templates that ship pre-loaded all run locally. Consult Mode is opt-in for the frontier-escalation case.
The pricing is flat. License + electricity, not per-call. The economics work because we don't pay LLM costs per user — they do, and locally, that's electricity.
The deployment is single-binary. Install on a laptop, the product works. No vendor cloud to log into, no service to depend on.
The product looks the way it looks because of the bet. The bet is the product.
What we want from you
If you've been frustrated by the cost compounding of cloud-LLM-based agent platforms, the privacy posture, or the lack of customization, give us a week. Free Desktop tier, no card. 7 templates pre-loaded.
If we're right about the architectural shift, you'll see it in the first hour of usage. If we're wrong, the cost to find out is your time.
Either way, we'd rather you make the call based on running the product than reading our thesis about it.