Measure workload
Record task volume, active duration, peak concurrency, turns per Session, artifact size, and retention. These are capacity methods, not Provider price quotes:- Steady-state active work is approximately arrival rate times mean active duration; account separately for bursts and tails.
- Daily artifact growth is approximately completed Runs times mean exported bytes, plus events, inputs, history, and indexes.
- Model/Tool usage follows actual requests and upstream metering; retries and verification can also consume usage.
- Stopped compute may retain disks, snapshots, or associated endpoints.
A reproducible workload estimate
These are assumptions, not measurements or price quotes: accept work for eight hours daily at a smooth 60 Runs/hour; each Run executes for 120 seconds on average, captures 4 MiB of artifacts, and retains them for 30 days. Initially exclude failed attempts, authorization waits, startup, idle and cleanup overhead; add them back for real planning.
Ten simultaneous arrivals are not represented by an average of two. Approval waits can extend resource retention beyond effective execution. Measure arrival volume and distributions of queueing, startup, execution and confirmed cleanup time, including tails, before setting admission limits and capacity.
Add overhead without counting it twice
For a sensitivity estimate, assume every daily Run allocates a distinct resource, startup takes 30 seconds, each resource remains allocated for another 60 seconds after execution, and 10% of Runs need one additional 120-second attempt on that same resource. Assume these intervals do not overlap for a given resource:
This is occupancy under a hypothetical resource lifecycle, not a bill or a recommendation to retry failures. An unknown attempt requires reconciliation before deciding whether another attempt is safe. Additional allocation on retry would add its own startup and retention. Reuse across Runs changes this calculation: measure the union of allocated intervals per resource, then sum across resources; do not charge the same warm interval to every Run. Apply the Provider’s actual metering and rounding separately. Retained storage, model requests, Tools and the control plane still need their own usage records.
Default worker capacity
The current default startup creates oneExecutionWorker. Its loop claims a Run and waits for that execute call to return before claiming another. Database leases and fencing enable takeover and reject stale writes; they do not establish verified horizontal scaling for an arbitrary multiprocess deployment.
In the idealized example, one serial execution slot handles at most about 30 tasks/hour at 120 seconds each, before overhead. Sustained input of 60/hour therefore accumulates waiting work. This is an assumption-based capacity warning, not a benchmark. More Sessions alone do not prove higher throughput. One active Run per Session and simultaneous execution capacity for the whole deployment are separate limits. Multi-worker deployment requires validation of shared configuration, database, credential ingress, resource locking and failover.
Observe and respond
/v1/overview counts nonterminal active_runs. /v1/sessions/activity measures task activity and its categories can overlap. Neither measures CPU or bills. Fleet is a persisted snapshot capped at the latest 200 Sessions and cannot prove account-wide cleanup. Combine complete paginated records with Provider checks; see the monitoring contract.
This page does not quote changing unit prices. In your budget record, capture the official Provider/model pricing page, lookup date, account plan, region, billing unit and storage/network components, then apply measured usage. Do not add unlike units or exclude failed attempts and cleanup waits from usage.