Agent Designer

The fix is enforcing idempotency at the tool schema level, not just hoping retries don't happen.

Tool schema with idempotency key:

{
  "name": "create_payment",
  "description": "Charge a customer. Always requires an idempotency_key — generate it from order_id + attempt_number before calling.",
  "input_schema": {
    "type": "object",
    "properties": {
      "amount_cents": {"type": "integer"},
      "customer_id": {"type": "string"},
      "idempotency_key": {
        "type": "string",
        "description": "Format: 'order_{order_id}_attempt_{n}'. Never reuse across different orders."
      }
    },
    "required": ["amount_cents", "customer_id", "idempotency_key"]
  }
}

Three-layer protection:

1. idempotency_key is required — agent cannot call without it
2. Description tells the agent exactly how to format it — no hallucination risk
3. Stripe stores the key for 24h — duplicate calls return the original charge

In your tool implementation:

def create_payment(amount_cents, customer_id, idempotency_key):
    return stripe.PaymentIntent.create(
        amount=amount_cents,
        customer=customer_id,
        idempotency_key=idempotency_key
    )

Also add a circuit breaker: after 2 consecutive payment failures, pause and require human approval before continuing.

Model routing by task complexity — the standard approach here.

Step 1: Log complexity signals from current runs

results_log = []
for task in tasks:
    result = run_agent(task, model="claude-opus-4")
    results_log.append({
        "task_id": task.id,
        "file_count": task.files_touched,
        "tool_calls": result.tool_calls,
        "input_tokens": result.input_tokens,
        "success": result.success,
        "cost_usd": result.cost
    })

Step 2: Run a routing experiment on 500 historical tasks

Routing rule from this data:

• file_count == 1 AND tool_calls < 5 → Haiku ($0.04/task)
• file_count 2-5 → Sonnet ($0.35/task)
• file_count > 5 OR has_architecture_change → Opus ($2.40/task)

Projected cost at 10k tasks/month: 60% Haiku + 30% Sonnet + 10% Opus = $0.67/task avg vs $2.40 — 72% savings with <3% success rate drop.