Before I argue about what matters in agent infrastructure, I want to be blunt about what does not. The agent harness — the CLI that wraps the model and gives it tool access — is not a defensible layer. It is a commodity, and the sooner you accept that, the sooner you can build something that lasts.
Look at what real companies are doing. Spotify cycled from a homegrown loop to Goose, then to Aider, then to Claude Code. Stripe forked Goose and bolted their existing dev infra on top. Bitrise hand-rolled their own because they are a CI/CD platform and it was trivial. Nearly everyone else is converging on Claude Code or OpenCode with few exceptions. Augment bet heavily on owning this layer and is now in a difficult spot precisely because foundation model providers keep shipping their own CLIs that replicate the work.
The layers that are not commoditized — the ones where almost every company is building from scratch — are context, memory, and skills; orchestration and session state; and tool integrations that wire agents into internal systems. Survey the landscape and the pattern is stark: the harness column is almost entirely blue (bought or forked off the shelf), while context and orchestration are almost entirely orange (built in-house). There is no standard. There are very few vendors serving these layers well. And these are the layers that determine whether an agent system actually works.
I've argued elsewhere that skills are software, not markdown — and that distinction matters because the skill layer is where the actual differentiation lives. Wrapping a model is not the business. The orchestration, the context engineering, the deterministic checks that catch hallucinated types before they ship — that is the business.
If you are building a company on top of a harness you forked last quarter, you are building on sand. The harness will be free, fast, and forgettable in eighteen months. Build above it.
Sources
Related Essays
The Agent Harness Problem
Enterprise agents need layered interfaces, real software skills, and flexible platforms. The harness around the model matters more than the model.
Agents Are Software, and Software Needs a Factory
People talk about agent harnesses as if the harness is the interesting part. It is not. The interesting part is the factory — sandboxing, orchestration, persistence, model translation.
The Harness Layer Has No Moat
The agent harness — the loop that executes — is no longer a differentiator. The opportunity lives one layer up, in the abstractions an engineer Googles at 2 a.m. when the duct tape breaks.
Key takeaways
- The agent harness layer is fully commoditized. Companies switch between Claude Code, OpenCode, Goose, and Aider without friction.
- Context, memory, skills, orchestration, and tool integrations are the layers being built in-house at almost every enterprise.
- Wrapping a model is not a business. Solving the problems between the model and the work is where the company gets built.
FAQ
Which harnesses are companies actually using?
Most are converging on Claude Code or OpenCode, with a long tail using Goose, Aider, or hand-rolled internal harnesses. Spotify alone has cycled through several in twelve months — that level of churn tells you the layer is not defensible.
Why is the harness not a defensible product?
Foundation model providers ship their own CLIs, open-source clones replicate every meaningful feature within weeks, and the cost to migrate is a weekend of work. The harness is a force multiplier for an existing product, not a standalone offering.