Seven primitives that compose a fleet control plane. Each solves an organizational problem that loop engineering does not.
Job: Single source of truth for what agents exist.
- Agent ID, owner, version, status (active / paused / retired)
- Linked loop pattern(s) and skills
- Last reviewed date
Artifacts: FLEET-STATE.md, agents/registry.yaml, or platform workspace catalog.
Job: Define who the agent acts as.
| Mode | Behavior | Example |
|---|---|---|
| Claw | Fixed service credentials | @vendor-intake Slack bot |
| Assistant | Per-user OAuth | Personal Notion assistant |
Without this, audit fails at "with what authority?"
Job: Control who can clone, run, or edit each agent.
LangSmith Fleet levels:
- Can clone — fork to customize
- Can run — use without editing config
- Can edit — full configuration access
Job: One place for humans to approve, reject, or review agent actions across the fleet.
- Claw agents: editors see fleet-wide threads
- Assistant agents: private per-user inbox for sensitive work
Job: Cross-agent traces and searchable evidence.
Minimum: tool calls + decisions + principal + agent version.
Enterprise: retention, export, correlation across handoffs.
Job: Prevent runaway spend and attribute cost.
- Per-agent daily token caps
- Per-team monthly budgets
- Admission control when over cap
- Cheap triage loops before expensive sub-agents
Template: templates/fleet-budget.md
Job: Organization retains kill and recovery authority.
FLEET_PAUSE_ALLor platform kill switch- Autonomy tiers (F1–F3) per agent class
- Rollback to last known-good manifest
flowchart TD
REG[Registry] --> ID[Identity]
REG --> PERM[Permissions]
ID --> INBOX[Inbox]
PERM --> INBOX
INBOX --> AUD[Audit]
AUD --> ECO[Economics]
ECO --> SOV[Sovereign Control]
You can start with Registry + Permissions (F1) before full audit infrastructure.