Unless you sell software, stop building the harness.
Most large enterprises already have one in the building. Claude Code, Codex, GitHub Copilot. Your developers are writing code with them today. And now the instinct is to build a custom harness around them, or instead of them. Custom loops, custom tooling, custom everything.
I understand the instinct. The harness feels like the product.
For a bank, an insurer, a manufacturer, it is not.
These harnesses are getting fully customizable fast. Within a year they will plug into almost any workflow and handle almost any kind of work. Building a competing harness means building against the vendors' roadmap, with a fraction of their budget.
The hard problem is deployment
The hard problem is somewhere else. It is deployment.
Right now these tools live on developer laptops. An agent that lives on a laptop is a demo. It sleeps when the lid closes. The real engineering is running all of it like infrastructure, the vendor harnesses and the agents your own teams build. In containers, always on, no laptop babysitting them, connected to the internal network, and doing all of that safely.
Infrastructure is a promise about Tuesday at 2 AM. The agent runs on a schedule somebody owns, inside the network, under policy, whether or not any particular laptop happens to be open. It survives lid closes, vacations, and reorganizations. A demo needs a person in the chair. Infrastructure needs an owner. Most agent projects have the first and never hire the second, and that gap is the whole difference between a pilot and a platform.
There is a quieter benefit. Harnesses change fast, and a harness spread across fifty laptops is fifty slightly different setups drifting apart with every vendor release. On the ground, the version is pinned, tested, and rolled out like any other infrastructure change. When the vendor ships something better, the ground absorbs it once, and the work never notices.
The layers of the ground
That is a different discipline. Sandboxing. Agent identity. Network policy. Cost controls. Observability. These layers decide whether agents survive contact with the enterprise.
Sandboxing comes first because it bounds the damage. The agent runs inside a fence, with the tools and data its task requires and nothing else within reach. Whatever it gets wrong, it gets wrong inside the fence.
Agent identity is what makes the fence auditable. Every agent acts under its own identity, scoped to the task, so every action lands in a log against a name. I gave this layer its own essay, because attribution is where accountability starts.
Network policy is decided by humans in advance and enforced by the ground: which systems an agent may reach, which data it may touch, which actions require approval. The agent proposes. Policy disposes.
Cost controls turn the meter into a budget. Ungoverned, an agent's spend is a surprise that arrives at the end of the month. On the ground, every run is metered and cost per completed workflow is a number with an owner. I wrote about the economics of that in The Token Trap.
Observability is the layer that decides whether any of the others can be trusted. Every action logged in a form you can replay. What you cannot replay, you cannot audit. What you cannot audit, someone eventually switches off, usually right after the first incident.
One ground, every agent
The same ground carries all of it: the vendor harnesses your developers already use and the agents your own teams build. That is the point of building it once, properly. Each new agent stops being its own infrastructure project and becomes a workload the ground already knows how to hold.
It is also the build-vs-buy answer seen from the other side. The ground is made of things only you have: your network, your data, your policies, your compliance obligations. A vendor can rent you an excellent harness. The ground it stands on is yours, or it is nobody's. That argument, in full, is the flagship essay.
The vendors will keep improving the harness whether we build one or not. Nobody is coming to build our ground.
Build the ground.