looping_without_ece_observation_mode.py

# examples/looping_without_ece_observation_mode.py

from veloryn import execute, print_summary
import anthropic

client = anthropic.Anthropic()

TASK_ID = "claude-loop-observation"

# No max_cost → observation mode
MAX_COST = None

COST_MODEL = {
    "input": 0.000003,
    "output": 0.000015,
}

previous_output = ""

def run_task():
    global previous_output

    prompt = f"""
You are improving a customer support automation system for an e-commerce platform.

Previous version:
{previous_output}

Refine the system with:
- Better handling of edge cases (refund abuse, delayed shipping)
- Improved escalation logic to human agents
- More realistic customer scenarios
- Clear decision flows

Make it more detailed and production-ready.
"""

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=800,
        messages=[
            {
                "role": "user",
                "content": prompt
            }
        ],
    )

    output = response.content[0].text
    previous_output = output

    return {
        "output": output,
        "usage": {
            "input_tokens": response.usage.input_tokens,
            "output_tokens": response.usage.output_tokens,
        },
    }


# Fixed steps (unbounded execution)
for _ in range(15):
    execute(
        task_id=TASK_ID,
        run_task=run_task,
        input_data="customer support refinement",
        model="claude-sonnet-4-6",
        cost_model=COST_MODEL,
        max_cost=MAX_COST,  # observation mode
    )

print_summary(TASK_ID)