CONCEPT Cited by 1 source

Cost cap for agent pipelines

Definition

A set of layered spending controls that prevent unbounded LLM inference costs in production agent pipelines. Necessary because agent systems process variable volumes with variable per-item costs (multi-turn reasoning, tool calls), making cost unpredictable without explicit caps.

Layers (as deployed at Databricks)

1. Deterministic filtering — the cheapest control: ensure alerts matching known-benign patterns never reach an LLM at all.
2. Per-batch spend cap — a cost tracker accumulates estimated spend across each processing batch; if a configurable cap is hit, remaining items are recorded as "skipped" (not lost, just deferred).
3. Daily alert cap — a hard ceiling on total daily cost regardless of inbound volume.
4. Per-category tool-call budget — within each item, prevents runaway discovery loops where an LLM might keep querying for additional context indefinitely.

(Source: sources/2026-07-06-databricks-scaling-security-alert-triage)

Trade-off

Cost caps mean some work items may be skipped during volume spikes. This is an explicit, documented trade-off: bounded cost at the expense of completeness. Skipped items are recorded for later processing or human review.

Seen in

• sources/2026-07-06-databricks-scaling-security-alert-triage — three-layer cost controls on Databricks' security triage agents.