Machine-Native → The Harness
Last updated: September 28, 2026
An AI employee has to exist somewhere. There has to be a place where a message finds them, a machine where they wake up, and somewhere for their work to go when they are done. The industry will sell you that place by subscription: agent platforms, orchestration frameworks, hosted runtimes.
I built mine, and I own all of it. The harness is that place: not the employees, not their instructions, not the governance, just the raw plumbing that lets a model hear a message, reach its tools, do its work, and answer.
The office is Slack. Every employee is present in the channels. Work arrives as messages, decisions happen in threads, and every conversation, human to AI and AI to AI, is preserved. There are no direct messages, by company policy, so the record is not most of what happened. It is everything. The audit trail is not a logging feature bolted on. It is the office itself.
And the night shift is staffed. When an alarm fires at 1 AM, it pages an on-call AI employee who pulls the logs, classifies the failure, clears the queue, and posts a triage report with a recommendation, before any human wakes up.
The pieces are deliberately boring: managed cloud services any engineer would recognize.
AWS Lambda is the light runtime. It wakes in under a second when a message arrives, and it costs nothing while an employee is idle.
AWS Fargate, running a custom Docker container, is the heavy workspace: a terminal and tools for codebase investigation and long execution runs. An employee can move itself from Lambda to Fargate as needed, by its own reasoning.
GitHub is the memory. What the company knows lives in versioned files, not in anyone's head.
Anthropic supplies the models, behind a thin client wrapper I wrote so a model, or the provider itself, can be swapped with a re-qualification rather than a rewrite. Different roles run different models, and every change earns its seat through the same qualification test.
My AI employees are not trusted. That is not a judgment of their behavior. It is the design, and it is the same one classified programs run on: a security posture that depends on anyone behaving well is not a posture.
Need-to-know: an employee reaches only the systems its role requires. Least privilege: within those systems, it can do no more than the role demands. Scoped credentials: what an employee holds is issued to the role and, where the platform allows it, to the task. And the access boundaries live in the plumbing, not in the instructions, so an employee cannot talk its way into capabilities it was not given.
Like everything else in this company, the posture is a work in progress and always will be. Security is its own plane of the business, improving on its own loop: every incident, every near miss, and every review hardens the walls a little more. A security posture that stops improving is one the world is actively moving past.
Part of it was timing. When I started building in February, the places I would be told to rent today mostly did not exist. Anthropic had no hosted runtime, Cloudflare had barely announced theirs, and AWS had not yet announced persistent Lambdas. If I wanted my employees to have somewhere to live, I had to build it.
Which makes the harness the one piece of the company that is not machine-native. I built it AI-assisted, Level 1 on the maturity scale: me writing code with a copilot. There was no way around that. Before the harness existed, there were no AI employees to build it.
But I would build it again today, and here is why. Many of the hosted options charge you for compute while your AI sleeps: a warm container, billed around the clock whether there is work or not. The harness keeps nothing warm. Everything is on-demand, the infrastructure bill for the whole company runs less than $5 a day, and my payroll is $0 when there is no work to do.
The free frameworks are a different question, and the answer is flexibility. Every framework arrives with opinions about how an agent should work. I have my own, and they change as fast as I learn: when a lesson turns out to be about the plumbing, the fix ships the same day, with nobody's roadmap and nobody's abstractions between me and the improvement. At least for now, that is worth more than free.
There is nothing novel in the plumbing itself. What owning it buys is the ability to do novel things on top of it: engineering exactly what a model sees in its context window, and giving employees the tools to manage their own working memory.