Your AI Advantage Is in the Scaffolding, Not the Model

Six months ago your AI tooling stack was a pile of loose parts. Pick a model, wire up a harness, glue on some tools, hope it held. Now the parts are turning into products you can buy off the shelf. Routers pick your model. Factories run your agents. And the "build it yourself" default is getting harder to defend. Let me catch you up.

The deep cut

• The tooling layer is where the money moves now. Router, Warp Factories, and TrueForge all sell the scaffolding, not the model.
• Buy the harness when it stops being your edge. Warp targets teams without the resources to build cloud agents, memory, and evals themselves.
• The scarce thing is your team's attention, not tokens. Latent Space's Ryan Lopopolo called synchronous human attention the only truly scarce resource.

The part around the model got good

For two years the model was the whole story. That changed. Harness-Bench ran one model across 106 tasks in different harnesses and scores ran from 52.4 to 76.2. Same weights, a 23.8-point spread. Half the agent is the wrapper.

That is why the jump AI engineers noticed around Christmas was hard to name. It was not one new model. The model and the harness improved together, and their curves crossed. Claude Code hit roughly $1B ARR in six months by handing the model the loop at the moment it got reliable enough to run one.

For you, this means the "which model" debate is only half the decision. The tooling around it moves the number just as much.

Buying the scaffolding

Two companies just put the wrapper on the shelf. Ramp launched Router, a model routing service that switches between OpenAI, Anthropic, DeepSeek, and others through one API. You can send only hard problems to expensive models and watch token spend, latency, and cost on a dashboard. Ramp ran it internally for three years before selling it.

Warp went a layer up with Warp Factories, an out-of-the-box software factory. It ships the phases of software development, triage through verification, with agents you can run in the cloud, memory across them, and evals built in. CEO Zack Lloyd said the target is smaller teams without the resources to build all that plumbing themselves.

The build-versus-buy line just moved. A year ago you built this. Now you can rent it.

When building your own still pays

Buying is not always right. Stripe built a "minions" system inside its own codebase. Ramp built a background agent that watches its own code after deploy. Both chose to build because the tooling touched their core work.

The cost argument is getting concrete. TrueFoundry open-sourced TrueForge and matched Claude Managed Agents on a 14-task benchmark while using about 30% fewer tokens. Routing to a cheaper model cut cost around 75% with accuracy held. That kind of savings only shows up if you own the routing layer.

The tell is whether the harness is your edge. If it is generic plumbing, buy it. If it encodes how your team actually ships, keep it in house.

What no vendor can sell you

Here is the twist worth watching. As models absorb harness tricks, the scaffold shrinks. Anthropic's team deleted 80% of Claude Code's system prompt. What survives is the human side: permissions, trust, and when the agent is allowed to interrupt you.

That matters because the bottleneck moves. Warp's Lloyd said his own team automates 30 to 35% of tasks weekly, with a human still at the wheel for the rest. Berkeley's Dawn Song warns that errors accumulate across long sequences and that oversight has to be designed in, not bolted on later.

So even if you buy the router and the factory, the approval rules and the escalation logic are yours to build. Vendors give you the body. You still design how it talks to your people.

Three questions for your team

• Which parts of our AI stack are truly our edge, and which are generic plumbing a Router or a Warp Factory could replace next quarter?
• If we buy a routing layer, do we get the token spend, latency, and cost visibility we would need to catch a 75% savings, or are we flying blind?
• Who owns the approval and escalation rules, the part no model absorbs, and is that a real job on our team or an afterthought?