For about forty years, business software has done two jobs. It remembered what the business did, and it helped people do the work faster. Both were enormous. Neither did the work. A third kind of software is now starting to appear — software that operates a function of a business directly, with a human setting direction rather than driving every action. This is the case that it's a real new category, why the category matters more than any single product inside it, and what separates it from the very good tools that came before.
The system of record
The first era. ERP, accounting systems, the database, the CRM as a place to store customers. Software whose job is to remember — to be the authoritative record of what happened. This was revolutionary, because before it the state of a business lived in filing cabinets and in people's heads, and it didn't survive either of them leaving.
But a system of record is passive by design. It knows everything and does nothing. A human reads it, decides, and acts; the system writes down the result. Its genius is memory, not action. Salesforce did not sell to your customers. It remembered who they were.
The system of engagement
The second era — named, along with the first, by the analyst Geoffrey Moore around 2011. Software that meets people where the work happens and helps them do it: Slack, Notion, Figma, the entire SaaS explosion, and most recently the AI copilots. A system of engagement makes a human faster, more connected, more capable. This era is still going. Most software sold today is a system of engagement.
But look at what it still assumes: a human, doing the work, with the software in their hands. The copilot is the purest expression of it — it sits beside you and suggests, and you still type, still decide, still act. The better the tool, the faster the human. The output is still bounded by the hours that human is awake and at the keyboard.
What both eras share
Here is the thing that took forty years to become visible, because it was the water we swam in: a system of record and a system of engagement are both tools. A tool is operated by a person. Its value is a multiplier on human effort — sometimes a staggering multiplier, but always a multiplier on a human who is present, deciding, and acting.
Take the human away and the software goes still. The CRM stops being updated. The dashboard goes unread. The copilot has no one to assist. For four decades, "business software" has quietly meant "a tool a person uses." We stopped noticing the assumption because there was no alternative to notice it against.
That assumption is what's now breaking.
The system of operation
The difference is one of kind, not degree. A copilot that gets ten times better is still a copilot — a faster assistant to a present human. A system of operation with no one watching keeps operating: the function runs overnight, over the weekend, while the operator sleeps, because the operator's job was never to perform the steps. Their job was to set the goal, draw the limits, and review what came back.
This is not a thought experiment. It is what it looks like when a marketing function runs as software — content planned against the quarter's goals, drafted, published, measured, and reallocated toward what converts, with a human approving what matters and ignoring what doesn't. Or a finance function: transactions reconciled continuously instead of in a month-end scramble, anomalies flagged as they appear, a close package assembled and waiting for sign-off. Nobody is driving. The function is running.
Why this is a category and not a feature
You cannot reach a system of operation by adding an "autonomous mode" toggle to a system of engagement. Operating a function — as opposed to assisting with tasks — requires four things that tools have never had, because tools never needed them.
- Governance. Who is allowed to do what, on whose behalf, with what oversight. A tool doesn't ask permission because a human is the permission. A system that acts on its own has to carry the authority model inside it.
- Coordination. A function is many pieces of work acting as one. A system of operation has to make a population of agents behave like a single accountable function, including when something goes wrong.
- Memory. A function has to remember the business it operates — its customers, its constraints, what was tried and what happened. Not a chat history. The accumulated judgment that makes a veteran employee worth more than a new one.
- Boundaries. The system has to be structurally unable to exceed its mandate. Not "instructed not to" — unable to. Rails enforced in code, on the assumption that the reasoning will occasionally be wrong.
A tool with an autonomous mode and none of these is not a system of operation. It is an unsupervised intern holding your credentials. Those four properties are what make autonomy safe enough to be useful — and building them, not bolting an "auto" button onto an existing app, is the actual work of the category. We've written about that work in detail in harness engineering.
What it takes to be real
Categories get named long before they're real. The interesting question is always who ships the first instance that genuinely works — the first relational database, not the idea of one.
A system of operation is hard for a specific reason: the four properties above are hard, and on top of them the reasoning has to be good enough that a function run by software is a function run well, not a faster way to make mistakes at scale. What makes it possible now is the collision of two things that arrived within a few years of each other — frontier models good enough to do the reasoning a real function demands, and a discipline for wrapping them tightly enough that the reasoning is safe to act on.
Harnyss is built on the first and is, itself, the second. Agent reasoning runs on the Claude Agent SDK; the governed hierarchy, the memory, the coordination, and the rails around it are what we build. It runs functions, not tasks — marketing, sales, customer success, finance, legal, engineering, or any structure an operator defines. We did not name this era and we will not be the only ones building in it. What we're trying to do is ship the first version of it a real business can actually hand a function to.
What changes
Every era of software made the one before it look passive. The system of record made the filing cabinet look passive. The system of engagement made the system of record look passive. The system of operation makes the tool you are using right now — the dashboard you log into, the app you drive — look like exactly what it is: something waiting for a person to come operate it.
That's the shift. Not a better tool. A different relationship between software and the work. The tools don't disappear — you keep your record of customers, your documents, your copilots. But the center of gravity moves from software you operate to software that operates, and the people move from the keyboard to the wheel.
This is one of three pieces on that move. The companion essays look at what it means for the company itself and for the org chart.