Agent OS v2 Role Architecture
Explore this system
Step through curated paths without changing the source diagram.
Beat
Diagram guide
Explore this system
Inspecting compiled semantics
E ExportT ThemeS Style0 Reset+ Zoom in- Zoom outEsc Close
Find a node
/
No matching nodes
Route probe
Choose a start node
Pick two semantic nodes on the diagram
Semantic lens
Compare system roles
Choose up to two semantic kinds. One reveals its real traffic; two compare only direct authored relationships.
Choose a kind to inspect its nodes and touching relationships.
Semantic radar
Building overview
Semantic radar needs more MAP space.
How the agent team is organized
Agent OS v2 is the role architecture behind Chris's agent teams. It defines what each agent is allowed to do, how it is evaluated, and how roles compose into a runtime that does real work.
The structure:
- A role registry holds 19 role archetypes: core, department, and specialist.
- Six core roles are required: atlas, business-owner, qa-verifier, research-analyst, knowledge-steward, and platform-reliability.
- Nine department archetypes and four specialist archetypes cover the lanes of work.
- CONSTITUTION.json defines each role's authority and capability contract.
- SOUL.md encodes judgment and temperament; AGENTS.md holds shared non-negotiable rules.
- Business tenants (Brand75 and future packs) load as tenant packs via agent-assignments.json.
- An activation gate checks identity, tools, and evals before any role runs, then the runtime composes and invokes agents ephemerally.
This is the governance layer that lets an AI team run without a human watching every step.
Role model
- • Each role is a CONSTITUTION.json plus SOUL.md pair
- • Roles declare capabilities, not installed skills
Activation
- • Defining a role does not activate it
- • Brand75 pack status is defined-not-activated
- • Activation needs actor, scope, tools, authority, evals, and a kill switch
Protected
- • Roles cannot change their own constitution, SOUL, tool grants, or model policy
- • Those are version-controlled product changes