Token cap
1200
AgentOps ยท observability and reliability
agentops-trace-cost-guard
Retries and token use become operational risk when nobody can see cumulative cost or stop a failing tool.
Public · synthetic demoThe problem
Unbounded retries turn one flaky step into repeated cost and load.
Structured spans need to show attempts, tokens, logical latency, outcome, and cost. Budget and breaker decisions must be logged as policy events, not inferred after the bill.
Input
Scripted reliability scenarios with deterministic clocks and token counts.
budget_abort circuit_trip_fallback clean_run retry_recovery Pricing: illustrative static table Latency: logical ticks Model calls: none
The money shot
The hard cap stops spend before the over-budget step is applied.
Attempted cumulative tokens
1750
demo-medium cost
$0.008700
demo-small cost
$0.001815
The budget ledger retained 1150 used tokens and cost $0.003625 before abort. The retry scenario recovered on attempt 3. The breaker ended open after 3 consecutive failed steps and served fallback for the rejected call.
How it's verified
Trace and reliability logs tell the same story.
| Scenario | Measured outcome |
|---|---|
| budget_abort | 3 of 4 steps reached; aborted |
| circuit_trip_fallback | 4 of 4 steps handled; breaker open |
| clean_run | 3 of 3 steps completed; 1330 tokens |
| retry_recovery | 3 of 3 steps completed; 1030 tokens |
Honest limitations
These costs and latencies are instrumentation fixtures.
- The pricing table is made up for the sample and cannot estimate a provider bill.
- Token counts are supplied by scenarios, not measured by a tokenizer.
- Latency and backoff are logical ticks, not elapsed time.
- Traces are in memory with no persistence, concurrency, or resume path.
The guard reduces blast radius; it does not prove the underlying task succeeded. jigonyoo.com · Back to hub