Cost Controls for Long-Running AI Agent Workflows

The runtime limits needed before agent workflows create real model, sandbox, storage, and support costs.

Audience: Founders and platform owners pricing agent-powered workflows.

The problem

Agent cost is not only tokens. A run can consume sandbox minutes, storage, retries, egress, and support time.

Implementation path

Set per-run timeouts, max artifact size, concurrency limits, retry headers, and clear terminal states. Surface limit failures as product behavior, not mystery errors.

Tradeoffs and failure modes

Limits can reject edge cases, but they prevent one workflow from consuming the platform.

Runtime limits

max_run_seconds=1800
max_input_files=20
max_artifact_bytes=104857600
org_concurrency=3
retry_after_seconds=60

Run this on Argo