LG Energy Solution Builds ERCOT Power Bidding Optimization Agent Based on Amazon Bedrock AgentCore

LG Energy Solution has built an analytics agent based on Amazon Bedrock AgentCore to maximize operating revenue for its BESS (Battery Energy Storage System) assets in ERCOT, the largest wholesale power market in the United States. This system analyzes bidding history, BESS specifications, LMP (Locational Marginal Price), and settlement data through natural language queries to recommend price prediction models and revenue optimization strategies.

Architecture and Data Pipeline

The system consists of data pipeline, agent execution, state storage, observability, and security layers. Market power, weather, and grid data are collected via Amazon Glue, stored in Amazon S3, and analyzed through Amazon Athena. The agent execution flow involves user queries passing through Amazon Bedrock Guardrails to an LLM, which then calls MCP (Model Context Protocol) within the AgentCore Runtime.

MCP-Based Function Separation and Synthesizer

Core functions are separated into individual MCPs that process different types of information. MCPs responsible for internal data analysis, web search for external event collection, and complex regulatory interpretation are organically connected. When results from two or more MCPs are generated, the Synthesizer cross-interprets them to produce a reliable final answer.

Security and User Isolation Strategy

Considering the sensitivity of power trading data, Amazon Bedrock Guardrails have been applied. Personas are distinguished between general users and authorized users, with differential settings for Denied Topics and thresholds to block privilege escalation or exposure of internal architecture. Additionally, conversation history and agent internal working states (Checkpoint) are stored separately in S3 to isolate data between users and improve UI restoration efficiency.

Operational Monitoring and Evaluation

Execution logs and token usage are tracked via Amazon CloudWatch to design a service-unit billing structure. The quality of LLM answers and the accuracy of tool selection are continuously monitored using built-in evaluation metrics from AgentCore Evaluations (such as Correctness and Faithfulness).