Cortya Book a walkthrough

A named team of agents that runs your operations, and answers to you.

Personas with roles and rank. An orchestrator you brief in plain language. Memory that survives the week. And telemetry where every moving pixel traces back to an event you can name.

Runs in your environment, on your keys Consequential actions wait for a person Our telemetry layer cannot carry what your agents said

Four layers, and you can check every one of them.

Most AI platforms ask you to trust a demo. Each of these is backed by a published contract or an executable validator, so you can check the claim instead of believing it.

01 / PERSONAS

Named agents, not a chat box

Each persona has a role, a rank and a lane. It knows what it owns, who it reports to, and what it may not decide alone — because those are fields, not instructions in a prompt.

112 deployable personas across a full company org, split base and instance so an update never overwrites what a persona learned on your estate.

See the persona library

02 / ORCHESTRATION

You write the directive it answers to

You brief one orchestrating persona in your own words, and a prime directive you author governs the whole estate — including which decisions must stop and come to a person.

The directive contract has seven mandatory sections, and the orchestrator refuses to start on a directive that fails the checklist rather than proceeding on a guess.

Watch a decision reach a person

03 / MEMORY

It keeps what it learns, and drops what stopped being true

Lessons are captured where they happen, reconciled when they conflict, and promoted deliberately. Forgetting is a recorded act, not a silent gap.

One store format, three tiers, one explicit promotion step, and 15 named validator rules that run as code.

See memory consolidate

04 / SIGNAL

Every moving pixel traces to an event

One event contract describes everything the estate does. Anything that watches the estate — a dashboard, an auditor, your own tooling — reads the same stream.

Published and versioned, with 15 rules graded in CI and each one made to go red against a deliberately broken record before it was trusted.

And this page is itself a pure function of that stream: our build re-renders it and requires the bytes to match, and a second check requires every duration it renders to agree across every surface it appears on — so a sentence the stream does not support cannot ship, including a plausible one nobody notices.

Read the stream

How a day actually runs.

This is not a metaphor for the product. Each step below is a kind of event the estate emits, which is why the demo can show you the same sequence with nothing hand-drawn.

01

You write the directive

What the estate is for, what it may do alone, and what must always come to you. It is yours, in your words, and it is the thing every persona answers to.

02

Personas take their seats

Each one declares its role, rank, lane and tier as it starts, so the org you are running is a matter of record rather than a diagram someone drew.

persona.spawn
03

They work and brief each other

Work happens in the lanes. Personas hand context sideways, task downward and escalate upward, and every hand-off is an event with a direction.

message.send · execute_tool
04

One decision stops

A persona that will not decide alone raises the question with its options and a recommendation, and it climbs the ranks until it reaches you.

escalation.raise · escalation.hop
05

You answer, and it is recorded

Your choice goes on the record with who made it and whether it can be undone. Anything consequential is then staged, armed by a person, and run once.

escalation.resolve · decision.record · approval.execute
06

It learns from having asked

The lane captures what it learned, conflicts with what it already believed get reconciled, and the few things worth keeping are promoted to the whole estate.

lesson.capture · consolidation.run

Effective AI needs orchestration precisely because your systems are real.

An agent that can reach your systems is a security question, not a feature question. Each of these is enforced by code or by a schema — a property of the system, not a promise about our discipline.

The model never gets raw access

A model in the loop is given no tools that reach a real system. Every touch of your data goes through a sanctioned connector we wrote, on a default-deny allowlist.

Enforced in code, not by prompt instructions a model can talk itself out of.

Consequential actions are taken by people or by plain code

Reading, triaging and reporting run on their own. Anything that spends money or cannot be undone is staged, armed by a person, and executed once — and arming is not doing.

Every automated trigger records whether a threshold decided it or a model did.

Our telemetry cannot carry what your agents said

The attributes that would hold prompt and completion bodies are opt-in, off, and rejected by the contract's own validator. Not a logging policy — the record has no field for it.

A build that finds one refuses to publish the page.

Your keys and your data stay yours

The estate runs in your environment on your own model keys. For regulated work it runs entirely inside your compliance boundary, with nothing to send us.

There is no path by which your operational data reaches our infrastructure.

A decision climbing to a person.

Height is rank. Work happens low and wide, decisions travel up, and a person is the ceiling. The wide amber stretch is the estate holding still because nobody had answered yet — 15 min of it, of which the person's decision took 14 min. Two numbers, two measurements: what the estate did, and what the human did inside it.

executivesenior leadermanagerspecialistdomain SMEa personestate held still · 15 minNadia T2Owen T2Priya T3Theo T3Mira T3Cass T1invoke agentplanchatexecute toolretrievalembeddingssearch memoryfetch responsecreate agentinvoke agentPriya → Theo · peer · Share contextNadia → Owen · down · Request cost envelopePriya → Nadia · up · Request decision3 optionsanswered · option 2 (the recommendation)stagedarmedran13:0013:1513:3013:4514:0014:1514:30
a decision climbing to a person sent up sent down sent sideways advises the person, outranks nobody

What we have run, and what we have not.

The org model is real. The persona, orchestration and delivery layers were extracted from a working multi-app venture operation that runs this way every day — that is where the 112 personas, the band structure and the escalation discipline come from.

The memory and telemetry layers have never run in a real organization. They are specified, executable and graded in our CI, and no running estate has yet emitted one of their events. Everything you see on this site is a seeded synthetic replay and is labelled as one.

We would rather tell you that than have you discover it. The architecture and the delivery are what we are selling; production mileage is not, and any vendor claiming both this early is guessing at one of them.

We stand it up, and we keep it current.

An estate is not a signup. Someone has to learn how your operations actually work, write the directive with you, and stay while the first real decisions go through it. Then your operations change — and an estate nobody is maintaining starts quietly describing a company you used to be. That work is ours, and it is most of why this succeeds.

  1. We walk your operations — the workflows, the systems, and the decisions people are currently making by hand.
  2. We write the prime directive with you — what the estate is for, and the decisions that must always reach a person.
  3. We stand up the estate in your environment — personas, lanes, connectors to your own systems, on your keys.
  4. We stay for the first decisions — tuning what escalates, until the estate earns the autonomy you give it.
  5. We keep it current — new lanes, revised directives, personas retired and added, for as long as you run it. This is the part that does not end.

The closer

The best thing our estate did all day was nothing.

Every AI demo shows agents moving fast. Here the estate waited 15 min, because one decision was not its to make. Then a person answered — and the action ran, once.

One moment · 13:42 → 14:10 producer.kind synthetic

Raised · 13:42

executive Nadia senior leader manager Priya specialist domain SME

Priya would not decide this alone — 3 options, with a recommendation.

  • Option 1
  • Option 2, recommended and chosen
  • Option 3

The chips are empty because the stream is: it records how many options there were and which one was recommended, never what they said.

The decision moved from manager to executive.

Held · 13:45 → 14:00

15 min

Nothing moved.

13:45 14:00

Nadia stopped and put the question in front of a person, and the estate waited 15 min for an answer. No lane went on without it.

Answered · 13:59

A person chose option 2 after 14 min.

It was the estate’s own recommendation. Both the advice and the choice are in the stream, so that is measured rather than claimed.

  1. 14:01Written to the ledger, owned by Nadia and marked not reversible.
  2. 14:07A person armed it. Arming is not doing.
  3. 14:10The action ran, once, under the armed approval.

Armed by a person and executed by Nadia — two different hands, which is what makes it an approval.

Derived from 59 synthetic events, seed estate-day-1, against schema 1.0.0 — the same event contract the product emits. No event carries a content attribute, so nothing here is a transcript of anything an agent said. Read the whole day.

What an engagement looks like.

You subscribe to an implemented estate — the platform and the work that keeps it iterating as your operations change. Implementation is scoped against those operations, so the fee follows the scope and not a seat count.

WHERE WE START
One part of your operations that people are currently holding together by hand — the place where the cost of the status quo is already visible.
WHAT THE FIRST PHASE PRODUCES
A written prime directive, a running estate in your environment, and the first decisions going through it with your people watching.
WHAT THE SUBSCRIPTION COVERS
The running estate and the work of keeping it right: lanes added as each earns its place, autonomy widened against decisions you have already watched it make, directives revised when the operation moves under them.
HOW IT IS PRICED
Scope determines the implementation fee, because no two estates are the same shape — a single workflow and a multi-lane operation are different pieces of work. Tell us the scope and we will price it against that. There is no list price to look up and none is being kept from you.

Talk to us about your operations

Tell us what your team is doing by hand that they should not be. If you can include these, the first conversation gets much further:

  • The workflow you would hand over first
  • The systems it touches, and whether they have APIs
  • The decisions in it that must always stay with a person
  • Any constraint we should design around — regulated data, on-premise, procurement
Email hello@cortya.com

A person reads this mailbox and a person will reply — it is not a ticket queue. If it is easier, say when you are free and we will find a time.

See it against your own operations.

Walk a day in the estate first, then tell us which of your workflows should be the one it runs.