Agent Cost Control for ML Engineers
Your experiments, evals, and agent fleets spend on inference — and the bill scales with every parallel run. Cap it before it caps your GPU budget.
The ML problem
Eval fleets spawn hundreds of parallel agent runs — cheap per call, expensive in aggregate.
Retry loops on failing evals re-run inference continuously.
Model drift to pricier tiers changes the bill without a code change.
What ML teams get
Per-run and per-fleet ceilings enforced deterministically.
Velocity limits that kill retry loops mid-run.
Category rules: inference, data, and compute budgets stay separate.
How to deploy
Call sipi.bot from the harness before each spend, or add the guard tool to your agent framework (LangChain, LangGraph, Pydantic AI, ...).
What you get
| ML concern | sipi.bot answer |
|---|---|
| Eval fleet cost | Shared daily ceiling |
| Retry loops | Velocity limit |
| Model drift | Per-agent cap regardless of tier |
| Experiments | Category-scoped budgets |
FAQ
Does it add overhead to eval runs?
About 5 ms per check — negligible next to inference latency.
Can I budget by experiment?
Yes — per-agent rules let each experiment have its own ceiling.
Does it work with my framework?
LangChain, LangGraph, Pydantic AI, SmolAgents, LlamaIndex, and plain HTTP all supported.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check