Skip to content

Agents & Tools

This page summarizes the main agents and MCP tools in AlphaSearch.

Agentic Energy Architecture

1. Forecasting Agents

Typical forecasting options include:

  • Random Forest
  • (Extensible) LSTM, transformer-based models, TimeGPT, etc.

The forecast agent:

  1. Picks a model (optionally with a recommendation)
  2. Generates next-day price trajectories
  3. Annotates low/high price zones for arbitrage opportunities

2. Optimization Agents

Available controllers include:

  • MILP Oracle
  • Exact optimization with all constraints

  • Best for benchmarking and rigorous evaluation

  • Heuristic (time-based / quantile-based)

  • Simple rules: charge at night, discharge at peak

  • Fast and interpretable

  • RL Agent

  • Learns a policy from simulated or historical experience

  • Good for complex, changing environments

  • LLM-based Controllers (Gemini / Ollama)

  • Use language-model reasoning with embedded qualitative rules

3. Reasoning Agent

The reasoning agent answers questions like:

  • “Why did you discharge at hour 10?”
  • “What is driving profits here?”
  • “How sensitive is this schedule to small forecast errors?”

It uses:

  • The last forecast and schedule
  • Confined plots (prices, power, SoC)
  • Domain-specific heuristics

to generate short explanations in natural language.

4. MCP Tools

AlphaSearch exposes many operations as MCP tools, including:

  • milp_solve – solve the daily MILP arbitrage problem
  • plot_schedule – produce candlestick + SoC plots and animations
  • forecast_prices – generate price forecasts
  • Reasoning tools to summarize and explain behavior

These tools can be orchestrated by CrewAI, called from notebooks, or integrated into other agentic workflows.