Agentic scheduling. It starts with the machines, not with a chatbot.

Many vendors are selling "agentic scheduling" this year. Look closely and some of it is a chatbot sitting on top of your ERP, answering questions about a plan it didn't make, about machines it has never seen.

An agent that can't see your machines is a chatbot with a calendar. It can talk about your schedule. It can't fix it.

What "agentic" actually means

An agent isn't something that chats. An agent sees, understands, decides, acts, and then checks whether it was right. Take away any one of those five and what's left is a demo.

Atherya does all five. And the loop starts at the source: at the machine.

1 · It sees: the machine, not the master data

Before Atherya plans a single order, it has learned how every machine in your plant behaves. That means its normal on each product and each compound, in each season. It knows the machine's real rhythm, its typical stops, and its health as the sensors tell it.

Out of that comes one number for every machine, always up to date: effective capacity. That's the speed actually observed, times the availability actually measured, times the health actually read.

• Per product, where the evidence exists. If a press has run that item enough times, Atherya plans with the speed that press has on that item. If it hasn't, Atherya falls back to the machine's general rhythm and says so on every line.
• Chronic defect or degradation, never confused. A machine that has always been a little slow is just like that, and it isn't punished twice. A machine heading towards a failure loses capacity today.
• Remaining life only when it is provable. It enters the plan only when its unit is certain. If in doubt it stays out, and the plan tells you why.
• No measurement, no number. If there are no measured cycles, there's no capacity. There's a reason instead.

Their agent reads the catalogue. Ours reads the machine.

2 · It understands: what can go where, and what is about to break

Compatibility learned from history. Atherya knows which product can go on which machine because it has seen your plant make it there. Not because someone filled in a table. It will never put an order on a machine that has never run it.

Maintenance inside the plan, not next to it.

• A machine with an open failure, or stopped with no declared end, gets no orders.
• When the brain recommends a maintenance window, the scheduler reserves it in the plan, and work that crosses it waits.

Your scheduler and your maintenance team live in two separate systems that never talk. In Atherya they're the same brain.

3 · It decides: simple, explainable, honest

This is where the show often starts: "multi-agent negotiation", "swarm optimisation", buzzwords stacked on top of numbers that are wrong.

Atherya does the opposite on purpose. Orders are taken by due date and priority. Each one goes to the compatible machine that finishes it first, given the work already queued. That's it. You can explain it to a shift leader in thirty seconds.

The value isn't in the choreography. It's in having the right capacity. A simple plan on true capacity beats a sophisticated plan on fake capacity, every time.

And the comparison with your plan is honest by construction. Your current plan is projected with the same capacities, the same speeds and the same queues as Atherya's. The difference measures only one thing: how much better you allocate when you know your machines. No forecasting trick, no two different maths.

An order with no known compatible machine, or no measured capacity, is declared not plannable, with the reason. No invented assignments, ever.

4 · It acts: without waiting for Monday

Atherya doesn't replan on a calendar. It replans when the factory changes:

• a line slows down;
• a bottleneck forms;
• a failure stops a machine;
• a machine comes back.

The brain notices the world has changed and reorganises the plan straight away, with your sign-off and recorded execution proof, with just enough spacing not to chase noise. Even the reason it replanned is written down.

Then it moves the work on its own board. Every new order is added, every closed order removed, every move made, and each of them goes into a journal that can only be appended to, with its reason. Cards move live, in front of whoever is watching.

Every view of your planning shows the brain's plan: the scheduling calendar and production planning, with the machine your ERP had chosen shown right next to it. Your people stop dragging cards around. They pin: an order that must not move stays where a person put it, and the brain plans around it.

People pin what must not move. Automatic replanning stays within your signed guardrails, with your sign-off and recorded proof.

5 · It earns the right to act on its own

Autonomy isn't a switch. It's a ladder, and every rung is earned:

1. Shadow plan. The brain plans next to you and touches nothing. Every divergence is recorded and can be checked against what really happened.
2. Proposals. The brain asks: "shall we do it?"
   • Only material moves are proposed, and only a few open at a time. Five ratified proposals are worth more than fifty ignored.
   • Every proposal carries its full reasoning. You ratify a reasoning, not a number.
   • Proposals gain confirmations run after run, and expire on their own when the divergence disappears. It never asks you to sign an old photograph.
   • A move you rejected doesn't come back.
   • If two people decide at the same time, the second one sees the true state. Nobody overwrites anybody.
3. Autonomy with guardrails. This is a human concession, signed by a name, and going back down is one click. The brain ratifies its own proposals only when all the guardrails pass:
   • a very material move;
   • confirmed across several runs;
   • within a daily budget;
   • never a critical order;
   • never towards a machine at risk or under maintenance.
   Every decision is signed: a person, or the brain.

6 · It closes the loop: into your ERP and your MES

A ratified decision can go straight back into your ERP and your MES. This is governed write-back, not a free-for-all:

• Only ratified decisions. Never a plan, never an open proposal, never an idea.
• Off by default. A concession per plant, signed by whoever turned it on.
• Prepared mode. It records exactly what it would have sent, at zero risk, before anything is switched on for real.
• A durable queue. The row exists before the call does, and one decision is never sent twice.
• Strict templated operations. No free commands, ever.
• Never fatal. Limited retries, then a declared failure. The decision stays valid and the brain keeps working.

7 · It checks itself: a write is not proof.

This is the verification step that completes our approach.

After every decision, Atherya goes back to your ERP and looks:

• Is the order really on the proposed machine?
• Did it finish on the original machine instead?
• Did it go to a third one?
• Did it never move at all?

A write never certifies itself.

From that comes the only number that matters: how many decisions were ratified and actually executed, and how many hours they confirmed. The brain's own decisions are counted separately from the people's.

The next rung of autonomy isn't unlocked by a sales pitch. It's unlocked by that scorecard.

Where this goes: the factory that schedules itself

Scheduling is the first job the brain takes off your hands. It won't be the last.

The same loop is extending to everything that today lives in separate systems and separate teams:

• orders and materials, so that a missing component moves the plan before it stops a line;
• shifts and people, so that the plan knows who is on the floor, not just which machine is free;
• energy, so that heavy loads land where power costs least;
• maintenance, already inside the plan today, so that every intervention is part of production and not an interruption of it.

It is one condition understood, and one coordinated response. A machine slows down, and capacity, delivery dates, maintenance and energy move together, decided by one brain instead of reassembled by five people across four systems.

That's the destination: end-to-end, autonomous production management. Every rung of the ladder takes us there, and every rung is measured. The one thing that never enters the loop is safety: physical protections and intervention permits stay with people, at every level.

Today factories run software. We're building the brain that runs the factory.

Their agent vs ours

The bottom line

Agentic scheduling isn't a chatbot that talks about your plan. It's a brain that knows your machines, makes the plan, moves the work, writes it into your systems, and checks that it really happened.

Their agent talks about your factory. Ours runs it.