Setting up Cost Control for AI Agents in Paperclip
AI agents without cost control are a source of unpredictable billing. A single poorly configured agent in an endless loop can spend hundreds of dollars in an hour. Paperclip has built-in cost control mechanisms—we customize them.
Levels of restrictions
Per-agent limits:
max_cost_usd per task. max_tokens_per_call and max_calls_per_task. When the limit is reached, automatic stop and escalation.
Per-organization daily/monthly budget: The total budget for the entire AI team. A warning is issued at 80% usage. At 100%, new tasks are not accepted until the period is reset or the budget is manually increased.
Cost allocation: Expenses are tracked by agent, role, and task type. Identification of ineffective agents (high expenses/low-quality results).
Cost optimization
Routing expensive tasks on GPT-4o, simple ones on GPT-4o-mini or Claude Haiku. Caching of repeated requests (prompt caching in Anthropic reduces the cost by 90% for the cached part). Batching of independent requests.
Reporting
Weekly cost report: top 5 agents by expenses, trends, and cost per output unit. Helps make decisions about configuration optimization.