---
title: "The Model Is the Easy Part Now"
description: "How local inference, the agent harness, and a coming OS rewrite are reshaping the surface area of computing."
doc_version: "1.0"
last_updated: "2026-05-12"
canonical: "https://modiqo.ai/blog/the-model-is-the-easy-part.txt"
---

# The Model Is the Easy Part Now

*From the Factory Floor — notes from inside the workshop where rote is being built.*

There's a moment in every technology cycle when the interesting layer moves. The model layer was that for three years. It isn't anymore. The interesting layer now sits one step below the agent and one step above the silicon — the runtime substrate, the harness that turns probabilistic text into deterministic work, and the operating-system primitives the whole edifice still rests on.

## The surface-area inversion
Inference is no longer a thing that happens behind an API. It happens wherever there is silicon — laptop, workstation, edge box, phone. The bottleneck shifted from FLOPS to memory bandwidth and VRAM, and consumer hardware finally serves it. A fast inference engine in a vacuum is a Ferrari in a muddy field. The roads are the harness.

## The harness is the glue
Unix virtualized hardware into `process`, `file`, `socket`. The harness is the same kind of project, one layer up: *session*, *harness*, *sandbox*, *control plane*. 65% of enterprise AI failures in production trace not to model reasoning but to harness defects. The bottleneck moved with the model commoditization.

## Three topologies
Local-first, local-with-spillover, and remote-first. They are not competing. Any serious system runs all three concurrently under one control plane.

## Coding is the meta-use-case
A coding agent produces every other agent. The harness primitives that get sanded down for code — sandboxing, diff review, rollback, parallel branch exploration — are the same primitives every domain ends up needing.

## The kernel rewrite
`process`, `file`, `lock`, `cgroup` were designed for a slow human caller. Now the workload is hundreds of concurrent agents per machine. The filesystem buckles, package managers race on lockfiles, git worktrees become the dominant primitive, and the GPU memory layer needs a scheduler.

## Deadlock, revisited
Coffman's four conditions — mutual exclusion, hold-and-wait, no preemption, circular wait — were a curiosity at human pace. At agent pace the probability of all four co-occurring is no longer rare. The new question is what prevention looks like when actor count is 100x what the original kernel designs assumed.

## Divergence and convergence
Agents fork-join at machine pace. Worktrees are the prototype. Every other point of contention — package envs, DB transactions, network namespaces, GPU memory — needs the same isolation primitive.

## The phase shift
CPU → memory → I/O → network → model → substrate. Each phase, the interesting layer moves. The model is now the fast part. The substrate is the story. The primitives aren't named yet.

— Chetan