Project 10 · FinOps · AI Agent Infrastructure

AgentLedger

Multi-Agent Cost Engine

An AWS-native metering layer that tracks token usage per AI agent in real time — giving engineering teams the granular cost visibility needed to scale from 1 to 50+ agents without losing financial control.

View on GitHub Explore Architecture ↓

The Challenge

AI agents are a black box
for finance teams.

As organizations scale AI agent fleets, the question shifts from "does it work?" to "what does it cost?" Without per-agent metering, cloud bills become opaque — and scaling becomes a financial gamble.

The Problem

Untrackable Cloud Spend

Organizations running multi-agent AI systems lack visibility into which specific agents or tasks are consuming the most tokens. Token costs aggregate into a single line item — making it impossible to identify waste, optimize performance, or prove ROI to stakeholders.

The Solution

Granular Metering Layer

AgentLedger intercepts every agent invocation through API Gateway, records token usage per agent ID in DynamoDB, and emits real-time cost metrics to CloudWatch — giving teams a live financial dashboard for their entire AI fleet with zero code changes to existing agents.

System Design

How AgentLedger Works

A fully serverless metering pipeline built on AWS-native services. Every agent call flows through the ledger — tracked, stored, and surfaced as actionable cost intelligence.

🤖

AI Agent

Invokes LLM with agent_id header

→

🔗

API Gateway

Routes & authenticates requests

→

⚡

Lambda

Extracts & normalizes token data

→

🗄️

DynamoDB

Persists per-agent usage records

→

📊

CloudWatch

Real-time cost metrics & alerts

agentledger/lambda_handler.py

# AgentLedger — Token metering interceptor
import boto3, json, time

dynamodb = boto3.resource('dynamodb')
cloudwatch = boto3.client('cloudwatch')
table = dynamodb.Table('AgentLedger')

def lambda_handler(event, context):
  agent_id = event['headers']['x-agent-id']
  tokens = event['body']['usage']['total_tokens']
  cost_usd = tokens * 0.000002

  # Write to DynamoDB ledger
  table.put_item(Item={
    'agent_id': agent_id,
    'timestamp': int(time.time()),
    'tokens': tokens,
    'cost_usd': str(cost_usd)
  })

  # Emit CloudWatch metric
  cloudwatch.put_metric_data(
    Namespace='AgentLedger',
    MetricData=[{'MetricName': 'TokenCost',
      'Dimensions': [{'Name': 'AgentId', 'Value': agent_id}],
      'Value': cost_usd, 'Unit': 'None'}]
  )
  return {'statusCode': 200, 'body': json.dumps({'metered': True})}

Technology

Built on AWS-Native Services

Every component is serverless, managed, and scales to zero — meaning the metering layer itself costs nothing at idle and pennies at scale.

API Gateway

Request routing & auth layer

AWS Lambda

Token extraction & normalization

Amazon DynamoDB

Per-agent usage ledger

CloudWatch Metrics

Real-time cost dashboards

Terraform

Infrastructure as Code

Python / Boto3

Lambda runtime & AWS SDK

Business Impact

Granular ROI tracking
at any scale.

AgentLedger transforms AI infrastructure from a cost center into a measurable, optimizable system — enabling confident scaling from 1 to 50+ agents.

💰

Per-Agent Cost Visibility

Know exactly which agent is spending what — down to the individual task and token. Eliminate the guesswork from AI infrastructure budgeting.

📈

Safe Scaling to 50+ Agents

Prove ROI before scaling. AgentLedger gives finance and engineering a shared source of truth, making the case for expanding AI fleets with hard data.

🔍

Waste Identification

Surface underperforming agents consuming disproportionate tokens. Optimize prompt engineering and model selection with real usage data, not estimates.