Place 04 / Great Hall

Room 05 / 06 / The House

Great Hall

AI agents: systems that do the work and show their working.

The Great Hall

Since 2026 most of what I build is AI agents: language models that take actions, not only answer.

I built a team of coding agents that splits a job, builds it in parallel and checks its own work. I also built a finance app whose assistant never invents a number, because every figure comes from the user's own records.

I learned that an agent is only as good as the checks around it. Every run here is measured, and every claim can be traced back.

Plate 05 / Lanes

the karakuri figures / 2026

Software Factory

A team of AI coding agents. A Lead splits the work into lanes, builders work side by side, and scripts check each lane before it is merged.

The team needed between 39% and 64% of the time one agent took alone.

Read the chapter
Plate 06 / Ledger

the ledger and the abacus / 2026

Finance Agent

A personal finance app that runs on a Mac and can never move money. It sorts bank transactions and answers questions in plain English.

The month's overview now appears in under a third of a second instead of eight.

Read the chapter

the audit log

Every agent action, written down

At PA Consulting, agent workflows are built in LangGraph. pi.dev is being trialled to open a ticket for each agent action and log it, so an auditor can see every step.

the star chart

Constellation

A tool that draws a Claude Code setup as a map: what loads every session, what is never used and what overlaps.