How our AI works

How our AI agent system works: roles, reviews and human approval

Nexus runs the operation at Sigma 7 Services. It's the AI agent system we built for our own work. A request goes to an agent with one job, like writing, building or testing. A second agent reviews it, and a person approves anything that spends money, publishes, deletes data or messages someone outside the company.

Abstract tiles joined by a red line that pauses at a gate before finishing.
AI agent orchestration

What is AI agent orchestration?

AI agent orchestration is the system that decides which agent does which job, in what order, with what checks, and when a person has to step in. One agent doing everything is hard to check. Several agents with narrow jobs, a review step between them and a person at the gates is easier to trust. The system around the agents is sometimes called an agent harness. Nexus is ours.

Who does what

The roles in Nexus

Each agent has one job and a short list of what it may do.

  • 01CoordinatorTakes in the request, breaks it into steps and hands each step to the right role. Can't do the work itself.
  • 02ResearcherGathers what the job needs: the brief, past decisions, search data. Can't change anything it reads.
  • 03WriterDrafts the words: pages, emails, reports. Can't publish them.
  • 04DesignerLays out pages and visuals to the house rules. Can't publish them.
  • 05BuilderWrites the code and builds the change on a private test copy. Can't put it live.
  • 06ReviewerReads the work against the request and the rules, and sends it back with notes when it falls short.
  • 07TesterRuns the checks: links, forms, page speed, accessibility, banned words. Passes the work or fails it.
How a request moves

How do AI agents work together on one request?

Every request becomes a written ticket with a goal and a definition of done. It moves in one direction, and each step leaves a written record.

  1. 01Request
  2. 02Plancoordinator
  3. 03Workresearcher · writer · designer · builder
  4. 04Reviewreviewer
  5. 05Testtester
  6. 06Approval gatea personSpend · Publish · Delete · Outside messages
  7. 07Done
  • Review → WorkSent back with notes
  • Test → WorkFailed check

Every step is written to the ticket

A request moves from intake to a role, through review and testing, to a human approval gate before anything goes live.
  1. 01Request. A person asks for something, in plain words.
  2. 02Plan. The coordinator writes the goal, the steps and what "done" means.
  3. 03Work. The right role does its step: research, writing, design or building.
  4. 04Review. A reviewer checks it. If it falls short, it goes back with notes.
  5. 05Test. A tester runs the checks. A failed check sends it back too.
  6. 06Approval gate. If the next step spends money, publishes, deletes data or messages someone outside, it stops and waits for a person.
  7. 07Done. The work is delivered with proof of what was checked, and the record stays.
The gates

Where does a person approve the work?

In Nexus, an AI agent can draft, build and test on its own, but it stops and waits for a person before anything is spent, published, deleted or sent outside the company.

  • Spending moneyNo agent can commit a dollar. A person decides.
  • PublishingWork is built and checked on a private copy first. Going live is a person's call.
  • DeletingNothing is removed without a person's say-so.
  • Messages to the outsideAgents draft them. A person sends them, or approves the send.
RolePlanRead and researchDraftBuild on a private copySend backRun checksSpendPublishDeleteMessage outside
CoordinatorYesNoNoNoNoNoWaits for a personWaits for a personWaits for a personWaits for a person
ResearcherNoYesNoNoNoNoWaits for a personWaits for a personWaits for a personWaits for a person
WriterNoNoYesNoNoNoWaits for a personWaits for a personWaits for a personWaits for a person
DesignerNoNoYesNoNoNoWaits for a personWaits for a personWaits for a personWaits for a person
BuilderNoNoNoYesNoNoWaits for a personWaits for a personWaits for a personWaits for a person
ReviewerNoNoNoNoYesNoWaits for a personWaits for a personWaits for a personWaits for a person
TesterNoNoNoNoYesYesWaits for a personWaits for a personWaits for a personWaits for a person
PersonNoNoNoNoNoNoYesYesYesYes
Who may do what: each role's allowed actions, and the four actions that always wait for a person.
When an agent is unsure

What happens when an agent is unsure or a check fails?

It stops and says so, in writing, on the ticket. It doesn't guess its way past a missing answer, and it doesn't hide a failure behind a workaround. A failed review or test sends the work back to the role that made it. If the question is one only a person can answer, the ticket waits for that person.

AI agent governance

What AI agent governance looks like in practice

Governance is the set of rules that keep agents inside their lane: what each role may do, which actions always need a person and a written trail of who did what. In Nexus those rules live in the system, not in a policy document nobody reads. Every step is recorded, and the approval gates can't be skipped by an agent.

This site

This website went through the same process

The pages you're reading were researched, written, designed, built, reviewed and tested by roles in Nexus. A person approved them before they went live.

For your AI integration

How this applies to your AI integration

The same pattern goes into the systems we build for clients: narrow jobs for each AI step, a review before anything reaches a customer and a person's approval at the points you choose. You decide where the gates go.

Nexus Edu, our product for small colleges and seminaries, shares the name and the team. Nexus Edu 

Questions

Questions about how our AI works

How does Sigma 7's AI agent system work?

A request becomes a ticket, goes to an agent with one job, gets reviewed and tested by other agents, and waits for a person's approval before anything is spent, published, deleted or sent outside the company.

What is AI agent orchestration?

The system that assigns jobs to AI agents, sets the order and the checks, and decides when a person has to step in.

Where does a person approve the work?

Before spending money, publishing, deleting data or sending a message outside the company.

What happens when an agent is unsure or a check fails?

It stops and says so on the ticket. Failed reviews and tests send the work back, and questions only a person can answer wait for that person.

Can you build this for my company?

Yes. The same pattern of narrow roles, reviews and approval gates is how we set up AI inside your systems. Tell us what you want AI to handle.

Want AI that waits for your say-so?

Tell us the work you'd hand off and where you'd want to approve it.

Talk to us