跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
AI 与知识库

token-budget-tracking

Track, optimize, and control token consumption across multi-agent systems. Covers budget allocation, real-time monitoring, cost attribution, per-agent limits, and proactive cost optimization for production LLM deployments.

cosmicstack-labs · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add cosmicstack-labs/mercury-agent-skills --skill token-budget-tracking

技能说明


name: token-budget-tracking description: 'Track, optimize, and control token consumption across multi-agent systems. Covers budget allocation, real-time monitoring, cost attribution, per-agent limits, and proactive cost optimization for production LLM deployments.' metadata: author: cosmicstack-labs version: 1.0.0 category: ai-ml tags: - token-budget - cost-optimization - token-tracking - budget-alerting - llm-costs - resource-management

Token Budget Tracking & Optimization

Overview

In production multi-agent systems, token costs are the new infrastructure bill — and they can spiral fast. An agent in a loop can burn through hundreds of dollars in minutes. This skill covers how to set budgets, track consumption in real time, attribute costs to specific agents and tasks, and optimize token usage without sacrificing quality.


Core Concepts

Token Cost Economics

ModelInput ($/1M tokens)Output ($/1M tokens)Cost per 100K tasks (4K avg)
GPT-4o$2.50$10.00~$1,250
Claude 3.5 Sonnet$3.00$15.00~$1,800
GPT-4o-mini$0.15$0.60~$75
Claude 3 Haiku$0.25$1.25~$150

A single runaway agent consuming 50K tokens per loop for 100 iterations = $50-$150 in minutes.

Budget Dimensions

DimensionWhat It TracksWhy It Matters
Per AgentTokens consumed by each agentIdentify expensive agents
Per TaskCost per completed taskMeasure ROI per task type
Per UserCost attributed to a user/sessionBill-back, abuse detection
Per ModelCost by LLM provider/modelModel selection decisions
Daily/WeeklyAggregate burn rateBudget forecasting
Per StepTokens per reasoning stepDetect inefficient reasoning

Step-by-Step Implementation

Step 1: Build a Token Counter

from dataclasses import dataclass, field
from collections import defaultdict
import time
import threading

@dataclass
class TokenUsage:
    prompt_tokens: int = 0
    completion_tokens: int = 0
    total_tokens: int = 0
    
    def __add__(self, other: "TokenUsage"):
        return TokenUsage(
            prompt_tokens=self.prompt_tokens + other.prompt_tokens,
            completion_tokens=self.completion_tokens + other.completion_tokens,
            total_tokens=self.total_tokens + other.total_tokens
        )

class TokenCounter:
    """Tracks token usage across all agents with attribution."""
    
    def __init__(self):
        self.usage: dict[str, dict[str, TokenUsage]] = defaultdict(
            lambda: defaultdict(TokenUsage)
        )
        self._lock = threading.Lock()
    
    def record(self, agent_name: str, task_id: str, 
               prompt_tokens: int, completion_tokens: int):
        """Record token usage for an agent-task pair."""
        with self._lock:
            usage = TokenUsage(
                prompt_tokens=prompt_tokens,
                completion_tokens=completion_tokens,
                total_tokens=prompt_tokens + completion_tokens
            )
            self.usage[agent_name][task_id] = usage
    
    def agent_total(self, agent_name: str) -> TokenUsage:
        """Get total tokens for an agent."""
        with self._lock:
            total = TokenUsage()
            for task_usage in self.usage[agent_name].values():
                total += task_usage
            return total
    
    def task_cost(self, agent_name: str, task_id: str,
                  input_rate: float, output_rate: float) -> float:
        """Calculate monetary cost for a specific task."""
        usage = self.usage[agent_name].get(task_id)
        if not usage:
            return 0.0
        return (
            usage.prompt_tokens * input_rate / 1_000_000 +
            usage.completion_tokens * output_rate / 1_000_000
        )
    
    def top_agents(self, n: int = 10) -> list[tuple[str, TokenUsage]]:
        """Get the n highest-consuming agents."""
        with self._lock:
            totals = [
                (agent, self.agent_total(agent))
                for agent in self.usage
            ]
            totals.sort(key=lambda x: x[1].total_tokens, reverse=True)
            return totals[:n]

Step 2: Implement Budget Enforcements

class TokenBudget:
    """Enforce per-agent and global token budgets."""
    
    def __init__(self, counter: TokenCounter):
        self.counter = counter
        self.agent_limits: dict[str, int] = {}  # agent -> max tokens
        self.period_limits: dict[str, int] = {}  # period -> max tokens
        
        # Current period tracking
        self.period_start = time.time()
        self.period_usage: dict[str, int] = defaultdict(int)
    
    def set_agent_limit(self, agent_name: str, max_tokens: int):
        """Set a per-agent token budget."""
        self.agent_limits[agent_name] = max_tokens
    
    def set_period_limit(self, period_name: str, max_tokens: int):
        """Set a global period budget (e.g., daily, weekly)."""
        self.period_limits[period_name] = max_tokens
    
    async def check_and_apply_budget(self, agent_name: str, 
                                       estimated_tokens: int) -> bool:
        """Check if this request would exceed any budget. Returns True if allowed."""
        
        # 1. Check per-agent limit
        agent_limit = self.agent_limits.get(agent_name)
        if agent_limit:
            current = self.counter.agent_total(agent_name).total_tokens
            if current + estimated_tokens > agent_limit:
                return False  # Agent budget exceeded
        
        # 2. Check period limits
        current_period = self._current_period()
        for period, limit in self.period_limits.items():
            if self.period_usage[period] + estimated_tokens > limit:
                return Fals
阅读完整来源 ↗