Königswasser: How Agentic Workflows Dissolve the Application

A gold application window dissolving in a flask of Königswasser

Gold resists almost everything. Fire, air, water and most acids leave it untouched. Alchemists found one liquid that dissolves it anyway, and they called it Königswasser: royal water, because it could dissolve the king of metals. Its secret is that it is a mix of two acids, and neither of them can dissolve gold on its own. One attacks the gold. The other binds what comes loose into a stable form, so the reaction keeps going instead of stalling.

I keep coming back to that image when I think about what agentic workflows are doing to enterprise software. For forty years, the application has been our gold: the solid, valuable unit in which software is built, sold, learned and used, and the place where people do their work. Agentic workflows are the first acid: they take that packaging apart. The second is an open, connected model of the whole enterprise, a world model that agents can work with. It holds what comes loose, so it doesn’t settle into new silos. Together they are the Königswasser, and the application is starting to dissolve.

Two acids dissolve the application: agentic workflows and an open world model built on OpenUSD

Planning the week, today

Take something that happens in almost every manufacturing company, every single week. On Thursday, the production planner has to decide what gets made next week: which orders, on which machines, by which people, with which materials.

She starts in the business system to see the open orders and their due dates. She checks the machine schedule in another tool and the shift plan in a third. She emails purchasing to ask whether the material for two urgent orders has arrived. That a machine is down for maintenance on Wednesday, she learns from a colleague in the hallway. In the end, it all comes together in a large spreadsheet that only she fully understands.

Each of these programs is good at its job. Each offers hundreds of buttons and menus, most of which she will never touch. And each knows only its own part of the story.

What really connects them is not software. It is the planner: her memory of where to look, which number to copy where, and whom to call. And when a rush order arrives on Tuesday, she starts over.

Today, the planner holds it all together: orders, email, machines, spreadsheet and shift plan

What dissolves, and what doesn’t

Many people picture AI agents, digital assistants that carry out tasks on their own, as robots that click through the same menus faster than we do. I don’t think that is what happens. Menus and buttons exist because people need them to find their way through hundreds of functions. An agent doesn’t need them.

So when agents arrive, a program is not simply automated. It comes apart into three pieces:

  • Fixed screens give way to artifacts. Menus, buttons and the jumping between programs lose their purpose. Instead, the agent creates what the moment needs: a small dashboard, a report, a comparison of options, a form with exactly the questions it needs answered. These artifacts are made for one task. They show people what they need to know, and they collect their guidance on how to proceed.

  • The expertise stays, and becomes callable. What makes a good program valuable is not its screens. It is the knowledge built into it: the rules for approving a change, the logic for scheduling a line, the way costs are calculated. That expertise stays where it is, but agents can call it directly, through standard connections such as MCP. A standard plug doesn’t mean every program speaks the same language, though. That is the job of the third piece.

  • Information and behaviour get connected. Locked inside one program, information helps only that program. Agents need to see how it all fits together: this order needs that machine, which needs that shift. So the pieces become connected in one shared model of the company, together with how things behave: how a machine runs, how long a changeover takes, what happens when a shift is one person short.

A program comes apart into three pieces: artifacts, expertise offered via MCP, and information and behaviour connected in the twin

No longer the place where people work, a program becomes a source of expertise and information that is called on when needed. The business system still owns the orders; the planning tool still owns its scheduling logic. Some screens stay, too: a designer still wants a CAD tool, an operator still wants the control panel. What disappears is the need for people to carry a workflow from one program to the next by hand. That is the dissolving: not of the value, and not of the systems, but of the packaging around them.

Information and behaviour gravitate to the digital twin

An agent that doesn’t know the company can only guess. Ask it to plan next week, and it will give you a plan that sounds right. Sounding right is not good enough when the answer decides what a whole plant does for five days.

What the agent needs is not more information but connected information. To plan a week, it has to know what exists and how it relates: which machines can make which products. What the current state is: which orders are due, who is on which shift, what is in stock. What it can do: move an order, swap a shift, shift maintenance. What would happen if it did. And who is allowed to decide. Today those answers exist only in pieces, spread across programs and people’s heads.

That is why I believe information and behaviour will gravitate toward a digital twin of the whole organisation: a living model of what the company makes, how it makes it, where, with whom, and how all of it behaves. Gravitating doesn’t mean everything moves in. The business system remains the authority on orders, the maintenance system on maintenance. A pipeline keeps the twin current, and the twin connects the pieces: what each one means, how it relates to the rest, and where the original lives. It exposes behaviour too, from simple rules to full simulations, so a plan can be tested before anyone commits to it. And the more agents work with it, the stronger the pull: every new task is easier when it starts from the same picture.

Information and behaviour gravitate to the digital twin through the OpenUSD pipeline

For the agent, the twin is its world model: its map of the company, and its test bench. For people, it is the place where they see what an agent suggests before they say yes. And it has to be open, because this is the second acid. An open, connected world model keeps what dissolves from settling into a new silo owned by one vendor. In our vision, its foundation is OpenUSD, an open standard for describing 3D worlds and everything attached to them, which started in film production. OpenUSD lets many sources be composed into one model: assets, their relationships, their context and their links to behaviour. It doesn’t have to hold every transaction; it holds the connections. That keeps the twin independent of any single vendor, and ready to connect one day with the twins of suppliers and customers.

The same week, dissolved

Now the same Thursday, in a company that has made this step.

The planner tells her agent: plan next week; the two orders for our biggest customer have priority. The agent starts in the twin: open orders, machines and their maintenance dates, who is on which shift, what material is in stock and what is still on its way. It also sees what it may change: orders can move, maintenance can shift, a second machine is qualified for the urgent part. Then it calls on the expertise of each program: the business system’s rules for order priorities, the planning tool’s scheduling logic, the delivery dates from purchasing. Before it shows the planner anything, it runs the plan through the twin’s model of the plant to see where it would jam.

A few minutes later, her agent shows her an artifact: a one-page view made just for this decision. It shows the proposed plan machine by machine, the two orders at risk and why, and asks her one question: move the maintenance to Friday, or run the urgent order on the older machine? She chooses, the agent updates the plan, and she approves. The approved changes are carried out by the systems that own them, through their own services, with a record of who decided what, and why.

She is not the only one working on it. The shift lead, the buyer and the plant manager work from the same plan in the same twin, but none of them sees the planner’s view. Each has their own agent, and each agent builds an artifact for what that person needs to know: for the shift lead, who works on which machine and when; for the buyer, which materials have to arrive by when; for the plant manager, whether the week holds and where the risks are. When the rush order arrives on Tuesday, the agent proposes a revised plan in minutes, and every artifact shows what the change means for the person looking at it.

The same week, dissolved: each role works with its own agent and artifact on the same digital twin

That only works with a layer between the people, the agents, the programs and the twin. Agents do not act on their own account: each one works on behalf of a person, with that person’s role and rights. The buyer’s agent may see supplier prices; the shift lead’s agent may not.

So an agent should never see or do more than the person it works for, in the twin and in every program it calls. That holds only if the same rights are enforced consistently all along the way: in the layer that delegates them, and in the systems that own the data. On top of that, something has to coordinate which agent acts when something changes, send each proposal to the right person for approval, and keep a record of every decision. Without it, many agents on one twin are chaos. With it, they are a team.

She did not open a single menu. Every application was involved, and none of them was visited.

Every agent acts for a person, with that person's rights: people, artifacts, agents, access and orchestration, the digital twin, the OpenUSD pipeline and the programs with their MCP services

What remains in the solution

Gold dissolved in Königswasser is not lost. It is still there, only in a different form, ready to be shaped into something new. The same is true for our applications: their expertise and their records remain. What dissolves is the application as the place where work happens. Work becomes independent of any single application, grounded in a shared, executable understanding of the company.

For leaders, I see three things worth starting now:

  1. Build your world model. Make clear how your products, production, projects and suppliers depend on each other, so agents have a map to work with. Build it on open standards like OpenUSD, so it stays yours.

  2. Build the digital twin pipeline. Connect your systems to the twin without copying everything into it: the twin composes the context, the systems keep the records. Agents test possible changes against this model; approved actions run through the systems that own them. Decide early who approves what, and how every decision is recorded. Trust grows step by step.

  3. Ask your software vendors how their programs will work with agents. Favour open, standard connections over closed ones.

Put together, the picture has three parts:

  • The world model: an OpenUSD-based twin that connects context, state, relationships and behaviour, kept current by a pipeline from the systems of record.

  • Capabilities: the expertise, operations and simulations of each program, offered to agents through standard interfaces such as MCP.

  • Governance and orchestration: the trusted layer that connects people, agents, capabilities and the world model.

This is where SyncTwin comes in. The SyncTwin Asset Registry is the foundation we build on. Today it holds the twin as versioned OpenUSD assets that any OpenUSD application can open, controls access by role down to individual parts of the twin, and notifies agents and pipelines when something changes. On this base, we and our partners build each company’s twin pipeline, with SyncTwin’s expertise in industrial twins and OpenUSD. Next comes the governance layer for agentic workflows: approvals, the coordination of agents and a record of every decision. The twin stays open, the programs keep their jobs, and people stay in charge.

The application had a good forty years. What replaces it will not take forty years: AI gets markedly more capable every few months, and each step makes the next one faster. The gold is still there, in the solution. What matters now is who shapes it first.