AI infrastructure · Internal system
Agent orchestration
A working system for coordinating AI agent sessions across more than 200 specialized roles, each with its own memory, event log, and handbook.
- Role
- Design & development
- Period
- 2025 to present
- Status
- In daily use
- Platform
- macOS
- Built with
- Python
- Swift
- SQLite
- JSONL event logs
- macOS Accessibility API
- Headless Chrome
- 200+
- role residences, each with its own records
- 1
- append-only event log per role
- 1,900+
- small command-line tools in the catalog
01
The problem
Running many AI agent sessions creates the problems a team has: work gets lost between sessions, two sessions edit the same file, and nobody remembers why a decision was made. This system gives each specialized role a residence with its own records.
How work moves
02
What each role keeps
- A handbook that points to its guides and documents.
- A SQLite knowledge base for procedures and file references.
- An append-only event log of what happened.
- Injected memory for rules that must reach every session.
03
Parts I built
- Switchboard, a Swift app and command-line tool that carries messages between about 200 roles.
- A Swift tool that drives desktop AI apps through the macOS Accessibility API in the background, without taking focus or using the clipboard.
- A visual diff tool that renders a reference page and a candidate in headless Chrome with fonts and color pinned, and fails on any difference.
- A catalog of more than 1,900 small command-line tools.
04
Why the records matter
The records make the work auditable. When something breaks, the event log shows what happened and the knowledge base shows what the role believed at the time. Lessons move into shared memory only after they are verified.
Scope & limitations
Internal. There is no public demo, and internal names are left out.