01 - Mental Model

Mental Model

Son-of-Anton is a delivery workflow for using AI heavily without surrendering authorship. It exists because most AI-assisted development drifts into one of two failure modes:

babysitting the agent so closely that it never builds momentum
letting the agent produce a giant diff that nobody can confidently review

The workflow says: keep the human judgment where it matters, and automate the work between those judgment points.

The Three Gates

There are three developer control points:

Plan - approve the what and why.
Decompose - approve the how, as a stack of thin tickets.
Closeout - approve that the delivered stack is done enough to merge.

Everything between those gates is orchestration. Worktrees, branch naming, red/green/refactor, review artifacts, PR creation, review polling, and state persistence are machinery in service of those gates.

The Key Sentence

Slice, review gate, explicit advance.

That is the product. Stacked PRs are useful because they make slices reviewable and ordered, but they are not the point by themselves.

What The Human Owns

The human owns:

product direction
scope trade-offs
ticket approval
final acceptance
decisions to defer, patch, or stop

The human does not need to manually shepherd every command once the approved slice is clear.

What The Agent Owns

The agent owns bounded execution:

reading the handoff
writing the failing test when required
implementing the smallest passing behavior
verifying locally
preparing adversarial review
patching prudent findings
opening and updating the PR
recording what happened in durable files

The agent is allowed to move, but only inside the approved boundaries.

Why Durable Files Matter

Chat memory is not a system of record. Son-of-Anton writes the important state to disk:

product plans
implementation plans
ticket docs
handoff docs
review prompts, reports, and ledgers
orchestrator state files

This means a new session can resume without pretending the agent remembers everything.