Amybuilds

Intent as an open standard.
Readable, checkable, portable, runnable.

AIM, the Agentic Intent Model, describes what a system should do: goals, roles, processes, decisions and rules, connected into one graph. The system can be software, a business process, a project or an entire organization. The description is readable text, independent of the technology that later realizes it.

  • Open standard, MIT license
  • Readable text, no closed format
  • Independent of technology
# Purchasing

## Requirements
- **APPROVAL** — Orders over €2,500
  are approved by the head of operations.

## Persona: HeadOfOperations
### Role
- Approves larger orders.

## Operation: Approve
### Summary
Approves a larger order.
### Authz
- `purchasing#Persona:HeadOfOperations`
### Ensures
- The order counts as approved
  — [satisfies](aim:purchasing#Requirements[APPROVAL])

Amy writes this down in conversation.

Readable.
Fifteen building blocks, each answering one question.

In notes and tickets, data fields, approval rules, views and deadlines blur together. AIM separates them into facets. Each can be read, reviewed and changed on its own.

Record

What information exists?

An order, an expense claim, a customer profile – and who owns them.

Operation

What is done or decided?

An action with its preconditions, permission and the outcome that holds afterwards.

Process

In what order?

The steps of a process, branches after decisions and what happens on errors.

Persona

Which person is involved?

A person, or a role a person holds: which views they open and which tasks they take on.

View

Who sees what?

A view with its information and the actions available there.

Event

What happened?

A recorded event like “claim approved” that other steps react to.

Trigger

What starts on its own?

A schedule, a deadline or an outside signal that sets work in motion.

Capability

What must the work be able to do?

A function such as “create a PDF”, named for what it does, never for a vendor.

Style

How should it look?

Colors, type and layout, in plain words or exact values.

State

Which status changes are allowed?

The life of a case, from draft to done, with no skipped steps.

Knowledge

What needs looking up?

A handbook, a regulation, supplier rules: knowledge the work consults.

Organisation

Which group acts as one?

A company, a department, a team or a supplier: a group that works together and acts as one.

Project

Which goal has an end?

A goal with an end, such as the website relaunch by March. It holds work but does not do it.

Software

Which software is described?

A website, an app or a system with its own data, operations and views. It can also do a step on its own.

Service

Which system is used?

A system such as SAP, Zoho or GitHub, reached through its interface. It can perform a step itself.

Checkable.
Every connection has a fixed meaning.

Reads, changes, triggers, satisfies: because every connection is typed, gaps, contradictions and impacts can be derived without interpreting the text.

Define it once.

Every relationship is stated in exactly one place. The way back is derived; nothing is maintained twice.

Find gaps.

Missing responsibilities, dead ends after decisions and conflicting permissions show up while modeling.

Meet requirements.

Every requirement is connected to the steps, views and deadlines that fulfill it.

See the impact.

When a rule changes, it can be derived which roles, views and processes are affected.

Portable.
Bound to no tool.

A model can be taken along, handed on and read by other tools.

Readable text.

A model is .aim files with headings and lists. They can be read, versioned and put in any repository.

Open standard.

The specification is public and MIT-licensed. Any tool may implement it.
View on GitHub

For any realization.

The same model serves as the working basis in Amy, as the specification for a coding agent, or as input for any other tool that understands AIM.

Runnable.
The model itself runs.

An AIM model is more than documentation. Amy runs processes directly from the model, with no separate application code. Every step names who does it: a person, Amy or a connected system.

  • Rules apply at every step
  • Decisions wait for the responsible person
  • Running cases keep their version

From quote to order

Illustrative example
New office equipment€4,800
  1. Quote received

    Lena submitted the document.

  2. Amy checks the details

    Items captured. Budget available.

  3. Jonas decides

    Orders over €2,500 need approval.

  4. Place the order

    Waiting for approval.

A human decision releases the next step.

The rules for this process

AI model
Allowed under the deployment’s model policy
Access
Read the quote, check the budget, place the order after approval
Human decision
Jonas approves orders over €2,500.
Connection
The existing purchasing system, through its configured API

What happened so far

  1. Quote submittedLena · 14 items, €4,800
  2. Checks completedAmy · the budget covers it
  3. Approval requestedJonas · required by the amount limit

This is what a whole model looks like.

TSV Seehausen: a sports club with departments, a member app, a tournament project and the annual fee run in one model.

An example model from the gallery, live and read-only.See all examples

Write AIM
without learning AIM.

Amy builds the model in conversation. Anyone who’d rather write it by hand finds the specification on GitHub.

Tell Amy how the work runs