Agent Burst is the per-agent-type bounded queue. Each tenant gets a per-tenant share of each agent's capacity. During a tenant's incident, that tenant's tasks burst above the share, but never to the point of starving other tenants. The platform stays responsive even during bad days.
Each tenant gets a guaranteed share (default = capacity / number of active tenants). When one tenant has spare capacity from its share, that capacity is offered to bursting tenants in FIFO order. When a tenant's share is needed back, in-flight tasks finish (no kill) and queued tasks for that tenant move to the front. No tenant ever starves; no tenant ever monopolizes.
Each agent type has its own capacity. Cheap, fast agents (log-triager on Haiku) might have capacity 32. Expensive, slow agents (incident-summarizer on Opus) might have capacity 4. The page shows current capacity, current usage, and a recommendation when the queue is consistently bottlenecked on one agent type.
Each tenant can see its own utilization on the page: how often you bursted past your share, how often you queued, how often a queued task timed out. The data helps a tenant make capacity-vs-cost decisions for their own usage. Other tenants' usage is never visible, only your own.
When a tenant wants to pause its agent activity (deploy freeze, maintenance, security review), one click drains its share gracefully: in-flight tasks finish, no new tasks dispatch. Other tenants are unaffected. Resume is also one click.
Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.
Bursting is fine. Bursting that locks out another tenant's incident is not. The queue makes the difference automatic.