The operating system around it is.

Most organisations are still framing the opportunity too narrowly. Can we build an agent. Can it answer questions. Can it automate a task. Can it connect to our systems. Those are useful questions. They are not the decisive ones.

A correction

In essay 002 we argued that the enterprise battleground was the control plane: registries, gateways, identity, policy enforcement, observability. That was right, and the major platforms have since moved exactly there.

It does not go far enough. A control plane governs what an agent is permitted to do. It says nothing about whether the agent should exist, who answers for it on a bad day, or how it gets better without quietly drifting.

They need more than a control plane. They need an operating system.

The definition

A useful agent is not a prompt with tools attached.

It is part of a wider system, and that system has to answer three questions before the agent touches anything that matters: what it is, what it may touch, and what it leaves behind.

What it is

Its role

Scope

The job it holds, and just as importantly the jobs it does not. An agent without a bounded role becomes a general-purpose assistant wandering the business.

Its authority

Decision rights

What it may settle alone and what it may only prepare. We set out the full ladder in essay 003.

Its operational owner

Accountability

A named human who answers for it in production. Not a committee, not the vendor, not the team that built it and moved on.

What it may touch

Its data boundaries

Access

Which systems and records are in scope, which are closed, and what happens at the edge of that boundary rather than in the middle of it.

Its tool permissions

Reach

What it can actually operate. Read access and write access are different risks and should never be granted in the same motion.

Its approval points

Human gate

The specific actions that stop and wait. Designed in from the start, not bolted on after something goes wrong.

What it leaves behind

Its escalation rules

Uncertainty

Where a question goes when the agent cannot resolve it. An agent with no escalation path will guess, confidently.

Its evidence trail

Proof

What it saw, what it recommended, what it changed, what it cost. Enough to replay a decision months later and defend it.

Its improvement cycle

Learning

How it gets better from production evidence without drifting away from the behaviour it was approved for.

Without it

Different teams build different agents. Permissions are granted inconsistently. Memory and state become unclear. Actions are hard to audit. Risk ownership is vague. Nobody has a reliable view of what the agent estate is actually doing. That is the sprawl essay 002 described, and it arrives faster than most boards expect.

That is not transformation. That is uncontrolled automation with better branding.

Where to start

Not AI everywhere. One consequential workflow.

Essay 003 set out what every AI-enabled workflow needs. This is the order to build it in. The important number is the first one, because a programme that starts everywhere finishes nowhere.

01 · Pick one consequential workflow 02 · Define the business outcome 03 · Give the agent bounded authority 04 · Instrument the workflow 05 · Define approval and escalation 06 · Measure the result 07 · Improve from production evidence

Consequential is the operative word. A workflow nobody cares about will prove nothing when it works, and cost nothing when it fails. Pick the one where a measurable improvement would be noticed by the board, and where a mistake would be noticed by a customer. That tension is what forces the operating system to be real.

That is how AI moves from demo to capability.

Where Praxis stands

Our view is simple. Enterprise agents need a governed operating system that defines how they are deployed, authorised, observed, improved and held accountable for real work. Not a platform decision. An operating decision, made once per workflow and revisited as the estate grows.

The agent is only the visible surface. The operating system is where the value, trust and scale will come from.

Point of View ← All essays

Previous Governed autonomy →

Begin

Which workflow first?

One conversation, no pitch deck. Bring the workflow that matters most, and we will scope the operating system around it.

Book a consultation → See the methodology