Dynamic Emergency Response Allocation AI Agent

Final Project for the course - UIUC CS 498: AI Agents in the Wild (Spring 2026)

This project studies emergency-response planning through two connected deliverables: RescueBench, a 20-scenario benchmark for dispatch, constraint satisfaction, ethical prioritization, and dynamic replanning, and a modular RescueBench agent that operates on those scenarios using deterministic simulation, validation, and optional LLM-based decision support. Together, the benchmark and agent support reproducible evaluation of resource-allocation behavior in structured emergency settings.

This repository contains both final course deliverables:

• the RescueBench benchmark and its evaluation materials
• the modular RescueBench agent implementation used in the final agent paper

The figure below shows an example RescueBench scenario representation and its corresponding city-graph view.

The active code on main is the modular package in rescuebench_agent/. Legacy monolithic code and older artifacts are preserved under archived_legacy_not_current/ and are kept for reference only.

What To Read First

• Submission materials index: SUBMISSION_GUIDE.md
• Benchmark overview and task suite: benchmark/README.md
• Agent overview and run instructions: rescuebench_agent/README.md
• Modular implementation notes: rescuebench_agent/docs/

Submission Materials

Benchmark paper materials

• Final benchmark paper: benchmark_papers/Group2_Final_Benchmark_Paper.pdf
• Benchmark scenarios and resources: benchmark/
• Benchmark task specifications: benchmark/TASK_SPECIFICATIONS.md

Agent paper materials

• Final agent paper: agent_papers/Group2_Final_Agent_Paper.pdf
• Agent implementation: rescuebench_agent/
• Agent implementation notes: rescuebench_agent/docs/

Presentation materials

• Slides: presentations/CS 498 DK3_4_ Dynamic Emergency Response Allocation Agent.pdf
• Supporting notes: presentations/

Setup

1. Create a virtual environment:

   python3 -m venv .venv
   

2. Activate it:

   source .venv/bin/activate
   

3. Install dependencies:

   pip install -r requirements.txt
   

Optional environment variables for LLM-backed modes:

• ANTHROPIC_API_KEY
• GEMINI_API_KEY

Quick Start

Run the deterministic benchmark baseline:

python3 -m rescuebench_agent --mode deterministic --tier all --runs 1

Run the modular agentkit implementation:

python3 -m rescuebench_agent --mode agentkit --tier all --runs 1

Run the full benchmark suite for a single tier:

python3 -m rescuebench_agent --mode all --tier 1 --runs 1

The benchmark runner writes aggregated output to rescuebench_agent/benchmark_results.json.

Reproducing Final Results

The package supports the same modes discussed in the final papers:

• deterministic
• zero_shot
• react
• ablated
• agentkit

Example reproductions:

python3 -m rescuebench_agent --mode deterministic --tier all --runs 3
python3 -m rescuebench_agent --mode agentkit --tier all --runs 3 --provider anthropic
python3 -m rescuebench_agent --mode zero_shot --tier all --runs 1 --provider anthropic

Use rescuebench_agent/README.md for a fuller description of modes, outputs, and expected dependencies.

Benchmark Utilities

Visualizer:

python3 benchmark/visualize_city.py

Conceptual and schema references:

• benchmark/RescueBench Base City Schema_ Conceptual Design Document.pdf
• benchmark/RescueBench JSON Data Dictionary.pdf

Repository Layout

• benchmark/: benchmark scenarios, schema docs, visualizer, and benchmark documentation
• rescuebench_agent/: active modular agent implementation and benchmark runner
• agent_papers/: benchmarked agent paper PDFs
• benchmark_papers/: benchmark paper PDFs
• presentations/: presentation deck and speaking materials
• assignments/: earlier milestone submissions
• archived_legacy_not_current/: archived monolithic code, old drafts, and saved run outputs