How it works — agent fleets
Four tiers. One governance ceiling.
Agent fleets are not a side channel around the pipeline. Commander → Orchestrator → Task → Tool inherit the same classification, budget, and audit controls that apply to your people — with a fleet-locked ceiling that subordinates cannot raise.
Tier 1
Fleet Commander
Sets the locked classification ceiling and budget envelope for the fleet. Activation freezes the ceiling — changes require retire-and-recreate.
Tier 2
Orchestrator
Decomposes work into Task Agents. Every subordinate LLM call still enters the gateway pipeline with fleet context headers.
Tier 3
Task Agent
Executes scoped work and can join collaborative sessions as an Automated Participant — visible to supervisors, governed like any other actor.
Tier 4
Tool Agent
Narrowest tier: tool and retrieval actions only. Content above the fleet ceiling is refused before the model is reached.
Classification
A Task or Tool action above the fleet ceiling is blocked with the same refusal path as a human prompt — audit row plus optional GRC webhook.
Budget
Fleet runs spend against a cost-code envelope on the fail-closed budget path. Exhausted allocation or Redis outage returns BUDGET_BLOCKED.
Evidence
Fleet refuses emit gateway.fleet.refused and can push fleet_policy_violation into your GRC destinations — not a separate agent identity plane.
Escalate-to-human uses the same HITL approval path as standalone agents. Supervisors review fleet activity from existing conversation and admin surfaces — no parallel identity system.
See the sixteen-step governance pipeline