ERC3 Agents

Agents for the ERC3: AI Agents in Action competition.

Quick Start

# Set environment variables
export ERC3_API_KEY=key-...       # Get from https://erc.timetoact-group.at/
export ANTHROPIC_API_KEY=...      # For Claude agents
export OPENAI_API_KEY=sk-...      # For SGR agents

# Run production agent (103 tasks, 5 parallel workers)
cd claude-agent-erc3-prod
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
./run.sh parallel 5

Agents

Agent	Benchmark	Architecture	Score
claude-agent-erc3-prod	erc3-prod	Anthropic Claude + Tool Use	100%
sgr-agent-store	store	OpenAI + Schema-Guided Reasoning	—
sgr-agent-erc3	erc3-dev	OpenAI + Schema-Guided Reasoning	—

Architecture

Claude Agent (claude-agent-erc3-prod)

Production agent with evolution system for iterative prompt improvement:

evolution/
  state.json          # Current version pointer
  v103/
    config.json       # Prompt, rules, examples, tool patches

Run commands:

./run.sh parallel 5      # Full run, 5 workers
./run.sh task t017       # Single task
./run.sh failed          # Show failures from last run
./run.sh version         # Current config version

SGR Agents (sgr-agent-*)

Schema-Guided Reasoning with Pydantic models and OpenAI structured outputs.

Results

Latest run (v103): 103/103 tasks (100%)

• Wall-clock time: 6.6 min (5 workers)
• Avg per task: 18.9 sec
• Tool calls: 596 total (5.8 per task)

Resources

• Competition Page
• ERC3 Platform
• Web UI — view sessions, tasks, logs