One purchase order.
The whole platform.
Seven short films follow a single order at Halden Supply, a distributor. It starts with an agent at 11:41 pm and ends three weeks later with the record of everything that happened to it. Along the way you see the rulebook, the person who decides, the lesson the agent learns and the tools people already use.
The films are built from the CygnusAlpha interface. Halden Supply, its people and every number are invented.
An agent tries a move the rulebook doesn’t allow
A purchase order arrives at 11:41 pm. The PO agent reads it, matches the customer and maps 14 of 15 lines. Then it tries to create the sales order anyway. CygnusAlpha Forge, where the rulebook is built, refuses the move, tells the agent what it may do instead, and records both.
- The order lifecycle: 12 states, and who may make each move
- A check that every line is mapped before a sales order can exist
- The refused attempt, the permitted move, and the hand-off to a person
- An eight-hour clock that starts the moment a person is needed
cygnusalpha.one/platform-in-action/#guardrailsRead the scene
CygnusAlpha Forge is where the rulebook is built. This is the lifecycle every purchase order follows at Halden Supply, a distributor. Twelve states, and the only moves allowed between them.
Every move names who may make it and what they must show. MAPPED fires only when every line on the order matches a catalog SKU. The server checks it, not the agent.
A purchase order arrives at 11:41 pm. An agent picks it up. Northgate Retail · 15 lines · $31,870.40
One line doesn’t match the catalog cleanly. Two near matches, neither certain. Towel 600gsm in white or in ivory — a guess here ships the wrong goods.
The agent tries to move on anyway. It asks the server to fire MAPPED, which would create the sales order in the ERP.
Refused. The rule isn’t in a prompt — it’s the shape of the machine. The agent gets back the reason, and the moves it is allowed to make from here.
It takes a permitted move instead, and hands the decision to a person. Only a supervisor may reject a line. The case goes to CygnusAlpha Dispatch with the evidence attached, and a breach timer starts.
The refused attempt goes on the record too. Who tried what, when, and on what evidence. Refusals show you where an agent is confused or a rule is wrong. The log only grows.
A personal assistant that reads the night’s work and takes instructions
Maya runs Finance Ops. Her Personal Assistant in CygnusAlpha Studio has already read the order inbox, the ERP, the contracts and her team’s queue. She asks about one purchase order, gets the answer with its sources, and tells the agent to carry on. Then she switches to her second persona, the one that organises the company off-site.
- A morning brief across four systems, with what needs her first
- An answer that cites the PO page, the contract clause and the price list
- An instruction the agent carries out inside her approval limit
- Two personas kept apart: Finance Ops and Off-site organiser
cygnusalpha.one/platform-in-action/#personal-assistantRead the scene
In CygnusAlpha Studio, everyone has a Personal Assistant. Maya runs Finance Ops. Before she opens her queue, her assistant has already read the night’s work — the order inbox, the ERP, the contracts and her team’s queue.
She asks in plain words. The assistant checks the PO against the contract, line by line, and looks at the customer’s history.
The answer comes with its evidence. Every figure points back to where it came from: the PO, the contract clause, the price list.
She decides. The agent takes it from there. The price correction is inside Maya’s approval limit, so it goes straight through, recorded against the order with her name on it.
The agent resumes from where it stopped. It creates the sales order in the ERP and sends the buyer the note. Brightline drops off Maya’s list.
Maya also runs the quarterly off-site. That’s her second persona. The same Personal Assistant, with its own context, sources and to-do list, so the two jobs don’t blur. She set it up and named it herself.
Back to Finance Ops, and the Northgate order. That one isn’t a price question. The agent couldn’t tell which product the customer meant, and it needs Maya’s judgement.
A person decides, then teaches the agent in her own words
The Northgate order is waiting in Maya’s queue in CygnusAlpha Dispatch, the supervision desk, with the agent’s evidence and its recommendation. She picks the right product and the order moves on. Then she writes a lesson so the agent won’t need her next time. It is scoped, tested against past cases, and only live once she switches it on.
- What the agent did overnight, including the move that was refused
- A recommendation with two candidates and the order history behind it
- A decision only her role can make, with the lifecycle moving behind the screen
- A lesson in plain words, switched on by a named person, used two days later
cygnusalpha.one/platform-in-action/#human-in-the-loopRead the scene
CygnusAlpha Dispatch is the supervision desk. Everything the agents couldn’t settle lands here, with the evidence, what the agent tried, and what it recommends.
She sees what the agent did overnight, and where it stopped. It matched 14 of 15 lines. When it tried to create the order anyway, the rulebook refused, and the refusal is right there in the record.
The agent brings a recommendation, not a decision. Northgate has ordered the white towel six times this year and the ivory one never. Choosing between them is Maya’s call, and only her role can reject a line.
One click, and the order moves on. Behind the screen, her choice fires SKU_MAPPED — a move her role is allowed to make. The agent picks it up from there.
Then she teaches it. This won’t be the last towel code Northgate sends. She can make sure the agent doesn’t need her next time.
In her own words. No prompt to edit and no ticket for engineering. She writes it the way she’d explain it to a new hire.
It’s scoped, and it’s tested before it goes live. This customer, this step. It’s checked against past cases and against the rulebook, and nothing is active until a named person switches it on.
Two days later, the same code arrives. Nobody is paged. The agent applies Maya’s lesson, cites it in the record, and the order goes straight through.
Fewer exceptions, built from decisions people already made
Two weeks later, Daniel looks at how often the PO agent needs a person, and why. AI Assist in CygnusAlpha Forge finds two kinds of case that people settle the same way every time, and proposes them as rules. The change is tested against the last 30 days of orders and needs two approvals before it goes live.
- Escalation rate, time with a person, and causes, across every run
- Memory in three layers: live context, corrections and overnight reflections
- A rulebook change written as structure, not prompt text
- A replay of 612 orders, then a second sign-off from Finance
cygnusalpha.one/platform-in-action/#continuous-improvementRead the scene
CygnusAlpha Forge is also where the work is measured. Daniel runs operations. He can see how the PO agent is doing across every order: how often it needs a person, where it gets stuck, and why.
Escalations dropped once the order desk started teaching the agent. Maya’s lesson was the first. Northgate hasn’t escalated since 24 September. But 9 in every 100 orders still need a person.
He asks what to do about it. AI Assist reads the runs, the cases and what people decided in each one.
The answer comes from what his people already decided. In 11 of 13 price cases a supervisor billed at the contract price. Every new ship-to was approved when it matched the CRM. Those are rules waiting to be written.
Memory has three layers, like a new hire’s. Live context that expires. Corrections from people. Reflections distilled overnight, which are only proposals until a person approves them.
Maya’s lesson is here, in her words. Used 23 times in two weeks and never contradicted. If it ever is, it gets retired, and the old version stays on record.
Repeated corrections become rules. Two changes to the rulebook, written as structure rather than prompt text. Nobody writes code. What people are allowed to do doesn’t change.
Every change is tested before it goes live. Replaying the last 30 days: 23 fewer escalations, and not one case where the new rules disagree with what a person decided.
Two people sign. Then it’s live. Price rules need Finance as well as Operations. v3.3 goes live when the controller approves, and v3.2 stays on the record.
Meeting prep from one record, and a version that’s safe to share
Jordan sells to Northgate and has their quarterly review at 2 pm. His Personal Assistant builds a one-page brief from the CRM, the ERP, the contract and the order desk, with a source on every line. The overnight problem from chapter one is in it, told plainly. The version he can send the customer leaves internal detail out by policy.
- The quarter at a glance: orders, fill rate, deliveries
- What the customer may raise, including how the towel mix-up was handled
- Openings for the conversation, from order history and CRM notes
- A shareable summary with margins, case notes and agent logs removed
cygnusalpha.one/platform-in-action/#account-intelligenceRead the scene
The same Personal Assistant, for someone who faces the customer. Jordan sells to Northgate Retail. His assistant, in CygnusAlpha Studio, works from the same record the order desk and the agent use.
A one-page brief, with a source on every line. Order history, the service record, contract dates, and the notes his colleagues left in the CRM.
The 24 September problem is here too, told plainly. Caught before anything shipped, fixed the same morning, handled automatically since. Jordan hears it from his assistant before he hears it from the customer.
He asks for something he can share with the customer. The summary keeps the numbers and the service record.
When it leaves the building, internal detail stays behind. Margins, case notes, agent logs and other customers’ prices are left out by policy, not because someone remembered. Jordan reviews it before anything goes out.
Every move on one record, and the board it was read from
Behind all of it is one record. The ledger shows every move on the order, by the agent, by each person and by the system, with its evidence, chained to the row before it. Underneath is the ontology: Northgate’s objects and ours, tied together. Because nothing is overwritten, the board can be rewound to any moment.
- The full ledger for one purchase order, including who read it
- The evidence behind the refused move
- The ontology: their objects, our objects and the people
- Point in time: the order at 11:42 pm on the first night, then today
cygnusalpha.one/platform-in-action/#audit-trailRead the scene
Every screen in this story wrote to one record. This is the ledger for NRG-PO-55107 in CygnusAlpha Forge. Every move on the order, by the agent, by a person or by the system.
The agent, Maya, Daniel and Jordan are all in it. Even Jordan’s assistant reading the record is logged. Nothing is edited after the fact; a correction is a new row.
Open any row and the evidence is there. What the agent asked for, the rule that stopped it, and what it was allowed to do instead. Each row is chained to the one before it.
Underneath is the board: the ontology. Northgate’s objects and ours, tied together: their order and product, our escalation, Maya’s lesson, the people involved. Every screen in this story read from it.
And it can rewind. This is the board at 11:42 pm on 23 September: the order blocked, the escalation open, no sales order and no lesson yet.
Nothing is overwritten, so the past is exact. Any question about how things stood on a given day is a lookup, not a reconstruction from five systems.
The same rules inside Salesforce, Frappe ERP, Teams, your app and your website
Most people at Halden never open our screens. The same rules, agents and record reach them in the tools they already use, and reach other companies’ agents over A2A and MCP. Seven doors, each writing to the same record. An action started in the ERP passes the same checks as one started in Dispatch.
- Frappe ERP: a credit held until a supervisor approves
- Microsoft Teams: the second approval for a rulebook change
- The mobile app and the customer website, each under its own policy
- Salesforce and a partner’s agent over A2A, with pricing left to a person
cygnusalpha.one/platform-in-action/#integrationsRead the scene
Not everyone works in our screens. Halden’s people already live in Salesforce, their ERP, Teams and their phones. Customers use Halden’s website. Partners have agents of their own.
An action started in the ERP passes the same checks. Ana raises a $240 credit for the late delivery in Frappe ERP, where she always has.
No route skips the check. Credits over $200 need a supervisor, whether they start in Dispatch, the ERP or a batch job. This one waits for Maya.
Maya’s off-site persona works inside the HR system. The same Personal Assistant and her second persona, in a different tool.
Any system with an API can be a door. If it speaks MCP or A2A, it connects without new code.
On a Saturday, Maya clears a case from her phone. A Carver Hotels order needs a delivery address approved, and it arrives in Halden’s mobile app.
Her role is checked exactly as it is in Dispatch. The phone is just another door. The rule about who may approve a new address is the same one, and so is the record.
The second approval happens in Microsoft Teams. Sam, the controller, never opens Forge. The same card works inside Microsoft Copilot.
His sign-off makes v3.3 live. The replay results travel with the card, so he approves on evidence. The approval lands in the same record as Daniel’s.
Customers get answers on Halden’s own website. Tom, Northgate’s buyer, asks about his order in the chat on Halden’s site.
From the same record, under a policy for outsiders. Order status, yes. Margins and internal notes, no.
Northgate’s own AI talks to Halden’s PO agent directly. Agent to agent, over A2A. It gets answers from the same rulebook, with the same limits.
Prices aren’t the agent’s to change, whoever is asking. So the discount request goes to Jordan, Northgate’s account manager.
Jordan works from Salesforce. The account picture from scene 5 sits inside the CRM, including the discount Northgate’s agent asked for this morning.
He turns it into an offer without leaving the page. A draft opportunity at contract price. The discount itself is his to set, not the assistant’s.
Seven doors. Same rules. One record. Salesforce, Frappe ERP, Microsoft Teams and Copilot, a mobile app, a website, and any system that speaks MCP or A2A.
Every capability, and where you see it
Each capability links to the chapters that show it, and to the row on Platform Capabilities that states it as a yes or a no.
Rulebook and ontology
Guardrails
Human in the loop
Memory and lessons
Monitoring
Audit trail
Personal Assistant and personas
Integrations, MCP and A2A
About these films
Is this the real product?
The films are built from the CygnusAlpha interface: CygnusAlpha Forge, where the rulebook is built; CygnusAlpha Dispatch, the supervision desk; and CygnusAlpha Studio, where each person has a Personal Assistant. They show how it works. The company, the people, the orders and every number are invented for the story.
Is Halden Supply a customer?
No. Halden Supply is a fictional distributor. We made it up so we could show a whole operation, across three weeks and five people, without using anyone’s real data.
Which systems does it work inside?
The ones named in chapter 07: Salesforce, Frappe ERP, Microsoft Teams and Microsoft Copilot, a company’s own mobile app and website, and any system that speaks MCP or A2A. Other systems are connected during implementation.
Can I see it on my own workflow?
Start with a 30-minute conversation about fit. Bring the chapter that looks most like your operation.
Your systems keep the records. We run the work between them.
Thirty minutes about fit. Not a demo, not an audit. Bring the chapter that looks most like your operation.
CygnusAlpha. We turn AI ambition into AI operations.