Skip to main content

Case study · Magnolia Orrery

An AI engineering team whose work can be checked.

AI agents do much of our building. Magnolia Orrery is how they work as a team: defined roles pinned to reviewed definitions, one identity per run, signed commits that name their ticket and session, and a record of every action they take.

IRIS (Integrated Reasoning & Intelligence Steward) is our planning assistant. The scoping page saves your answers to your own computer as a file, then offers a link to email them to us.

The problem

Fast work that nobody can audit.

An AI agent can write a great deal of code quickly. Without structure, nobody can say afterwards which agent did what, under which instructions, or whether the account that wrote a change also approved it.

A transcript is not a record. It ends with the session, and it does not say what was checked or what could not be verified.

What we built

Roles, identities and a record.

Defined roles

Ten specialist roles, from architect to validator, each with a written definition of what it may and must not do, and two coordinator roles that plan and oversee them.

Pinned definitions

Each role file is pinned by a fingerprint. An agent whose definition has changed since it was last reviewed is refused.

One identity per run

A run acts under exactly one role identity. The account that pushes a change cannot approve it, so nobody approves their own work.

A record of every action

Every tool call and dispatch is written as it happens to an append-only, hash-chained ledger, and every commit is signed and names its ticket, session and workspace.

How we built it

Written down, then enforced in code.

  1. 01

    Decide on the record

    Each rule starts as a numbered decision record that names what enforces it, or says that nothing does yet.

  2. 02

    Learn from the operator

    How we were asked to work is kept as numbered conventions, each saying how it is enforced.

  3. 03

    Check, do not trust

    Gates check the role fingerprints, the signatures and the record on every change, and a run confirms its own identity by probe rather than by assertion.

The result

The team, by the numbers.

Specialist roles
10
Architect, implementation, test, reviewer, validator and five more, each with written limits.
Role files pinned
12
The ten specialists and the two coordinators, each held to a fingerprint of its reviewed definition.
Decision records
88
Numbered and written down, each naming what enforces it.
Working conventions
14
How the operator asked us to work, each saying how it is enforced.

Measured 6 October 2026, from our main engineering repository as it stood that day.

Tell us what you need built.

Describe your systems, your reports and where the time goes on the IRIS scoping page. You keep a copy, and an engineer reads it before we talk.

Plan with IRIS