Architecture
A request enters through an interface, the Boss Agent plans and delegates, specialists do the work using tools and memory — and a governance layer sits around all of it. This describes the system as designed; parts of it are still being built and tested.
Nine things the Boss Agent does with every request.
- 01Understand the request
- 02Break down the task
- 03Assign specialist agents
- 04Manage dependencies
- 05Coordinate execution
- 06Collect results
- 07Review outputs
- 08Report progress
- 09Request approval for consequential actions
Every agent has a job description.
Each one is defined by four things: its responsibilities, its permissions, the policies it must follow, and how its performance is tracked. Not every agent operates fully autonomously in production.
| Agent | Responsibility | Defined by |
|---|---|---|
| Boss Agent | Understands requests, breaks them down, assigns work, reviews results and asks for approval. | role · permissions · policy · tracking |
| Planner | Breaks complex requests into structured plans. | role · permissions · policy · tracking |
| Researcher | Gathers and analyses information. | role · permissions · policy · tracking |
| Operator | Handles approved computer and workflow operations. | role · permissions · policy · tracking |
| Curator | Organises information and knowledge. | role · permissions · policy · tracking |
| Customer Support | Supports customer-service workflows. | role · permissions · policy · tracking |
| Finance | Assists with finance-related business workflows. | role · permissions · policy · tracking |
| Administration | Assists with administrative operations. | role · permissions · policy · tracking |
Persistent memory
A SQLite-based foundation for structured long-term memory: retrieving relevant context, keeping task history, organising knowledge, and giving continuity across workflows so agents don't start from zero every time.
Skills & playbooks
Reusable, text-based playbooks describe research procedures, operational procedures, quality checks and execution policies. The development environment holds 19 of them — working material for the agents, not 19 validated commercial capabilities.
