The agent that builds your agents.
Universal Constructor rebuilds your process as a simulated universe and breeds agents inside it, scoring each one on your metrics. Only the champion ever touches your business.
For processes with a measurable right answer
- Supplier invoicesaccuracy · cost per invoice
- Support triagerouting accuracy · first reply
- Lead qualificationconversion · time to first touch
- KYC document checksfalse accepts · review time
- IT ticket routingreassignments · time to resolve
- Order exceptionsrecovery rate · cost per case
Most agents are built once and left to drift. Your process changes, new edge cases appear, the models underneath get replaced, and the agent that fit on launch day slowly stops fitting.
One constructor. Endless generations.
Five steps, repeated for as long as the process exists. Each pass leaves you with an agent that scores better on the metrics you chose.
-
01 · Map
Map the process
Reads your procedures, tickets and logs, and interviews the people who run the process. Out comes a process map and the metrics that define success.
-
02 · Spawn
Spawn a population
Generates a set of agents that each take a different approach to the same job: other instructions, other tools, another step order or model.
-
03 · Simulate
Simulate
Builds a sandbox universe from your past cases and runs every variant through it. Nothing real is touched.
-
04 · Select
Select the champion
The best variant becomes the champion once you approve it. The weakest are retired. The rest keep competing as challengers.
-
05 · Adapt
Adapt
When results drift or the process changes, a new generation spawns from the current champion.
Evolution happens in the sandbox.
Your agents make their mistakes inside a simulated copy of your business, built from your own history. Only a proven champion crosses over.
- Beats the live championFitness 0.912 against 0.887
- Simulation fidelity94% match on 2,000 held-back cases
- Shadow modeMatched its score on 500 live cases, without acting
- Your approvalWaiting for the process owner
Nothing goes live until this box is ticked by a person.
Simulation speed
Evolution runs as fast as the sandbox can play cases, not as fast as real ones arrive.
Edge cases before they happen
The sandbox generates the rare, costly cases, like a supplier's new invoice format or a customer threatening legal action, so agents meet them in simulation first.
No blast radius
A failed variant fails inside the simulation. Your customers, suppliers and systems never see it.
Fidelity you can check
We replay past cases you held back and compare the simulated outcomes with what actually happened. You see that score before you rely on any result.
Only the fittest agents run your business.
Fitness is yours to define. Set what matters for this process and watch the ranking change.
| Agent | Gen | Accuracy | Cost / invoice | Handling time | Escalations | Fitness | Status |
|---|
Illustrative data for a supplier invoice process. Fitness is the weighted score of each metric, scaled across the population.
Nothing reaches your business until it has earned it.
The sandbox is the first line. These are the others.
Promotion needs your sign-off
A variant goes live only after it beats the champion in the sandbox, matches that result in shadow mode, and you approve it.
Every generation on record
Each agent's lineage, instructions, decisions and scores are logged. Roll back to any previous champion in one click.
Guardrails per agent
Set what each agent may access, how much it may spend, and when it must hand a case to a person.
Your data stays in the EU
Hosted in the EU and processed under GDPR. Your history trains your sandbox and nobody else's.
Your process changed in April. Your agent did too.
When live results drift from what the sandbox predicted, the constructor recalibrates the simulation on recent cases and breeds a new generation from the current champion.
Built once, or bred continuously.
| Question | Generic agent | Custom build | Universal Constructor |
|---|---|---|---|
| Fits your process | No | On launch day | And keeps fitting |
| Gets better over time | No | Only through new projects | With every generation |
| Measured on your KPIs | Rarely | Sometimes | Always. It is how agents are selected |
| Tested before going live | By the vendor | On a few sample cases | In a sandbox built from your history |
| When the process changes | It breaks | You start a new project | A new generation is bred |
Why “Universal Constructor”
In the late 1940s the mathematician John von Neumann designed, on paper, a machine that could build any machine from a description of it, including a copy of itself. He called it a universal constructor.
We borrowed the name and the idea. Our constructor builds the agents a process needs, tests them against each other, and keeps building better ones for as long as the process runs.
- 1948Von Neumann outlines self-reproducing automata at the Hixon Symposium in Pasadena.
- 1966Theory of Self-Reproducing Automata is published, completed and edited by Arthur W. Burks.
- 1995Renato Nobili and Umberto Pesavento run his design as software.
- 2026Universal Constructor starts building agents for businesses.
Become a design partner.
We are looking for a small group of companies to pilot one process each, from the first map to a live champion.
What you get
- One process evolved end to end, from mapping to a live champion
- A direct line to the founding team
- Preferential pricing once the pilot ends
What we ask
- Access to past cases and their outcomes for that process
- One person who owns the process on your side
- Thirty minutes of feedback each week
Questions
Will agents experiment on my customers?
No. Every variant is bred and tested in a sandbox built from your past cases. Only the champion goes live, and only after it matches its sandbox results in shadow mode and you approve it.
How realistic are the simulations?
We build them from your own history, then test them. We hold back a set of past cases, replay them in the sandbox, and compare the simulated outcomes with what actually happened. You see that fidelity score before you rely on any result.
What happens when reality drifts from the simulation?
We treat it as a signal. The constructor recalibrates the sandbox on recent cases and breeds a new generation from the current champion.
What data do you need?
Past cases for the process with their outcomes, your written procedures if you have them, and read access to the tools the process runs on. The more history you share, the more faithful the sandbox.
What exactly evolves?
An agent's instructions, the tools it calls, the order of its steps and the model behind each step. Each generation varies these and keeps what scores best.
Who owns the agents?
You do. The champions, their lineage and the sandbox built from your data belong to your company.
Does this replace my team?
It takes over the repetitive steps. Your team decides what fitness means, approves every promotion and handles the cases agents hand back.
How is it priced?
Per process. Design partners get preferential pricing for the first year.