System instructions, user tasks, files, tool results, and session history carried into a turn.
Jovethra / cost boundaries
Hard API quotas for agent workloads.
Agent usage is not one number. Input, output, cache, request rate, and concurrency each create a different operational risk. A hard limit makes the boundary visible before an open-ended bill is created.
Five limits, five questionsINPUTContext OUTPUTGeneration CACHEReuse Rate Concurrency Weighted cost
Input, output, cache, rate, and concurrency.
The visible response and model-generated decisions. Bound this explicitly for repeatable evaluation.
Repeated instructions and context can change the economics of long-running agent sessions.
Caps how quickly a client can begin requests in a short window.
Caps how many requests may be in flight at once, protecting both the account and the shared service.
Model selection can apply an additional cost boundary while raw category limits still remain in force.
Why separate limits matter
One total token number can hide a budget problem.
An agent can produce a short answer after processing a large repository, repeatedly reusing a long system context, or calling tools that add substantial input. Separate input, output, cache, rate, and concurrency controls make these dimensions inspectable rather than treating all tokens as interchangeable.
Published quotas are a planning boundary, not a guarantee that a particular prompt, model response, tool chain, or third-party dependency will succeed. Test the workflow with realistic but non-sensitive fixtures.
Practical calculationCount a representative task Multiply by the real cadence Set a stopping rule Estimate an agent workflow →
Estimate before you automate.
Measure input, output, and cache separately for one typical task.
Use daily or monthly task count, then leave headroom for retries and longer tasks.
When a limit is reached, decide whether the system should queue work, move to a new period, or require an explicit plan change.