Scry's congestion pricing turns an agent's retry into a purchase

A 429 sheds load because the server never runs the query. Request 101 is refused, and the client's timer decides when it comes back ^[6]. Scry answers the same overload with a higher quote, and an agent that accepts the quote gets its rows ^[1][3]. A retry into a spike is a purchase. The work that raised the price stays on the server, and the post argues that many agents retrying this way amplify congestion instead of damping it ^[8].

The timer is the part that does not fit. Standard exponential backoff waits one second, then two, then four ^[7]. That is four attempts inside the first seven seconds ^[15]. No step in that ladder reads the quote, so a spike that outlasts seven seconds gets billed four times ^[3].

Scry does not enforce an arbitration model, so the orchestrator implements one; the post's own example is ten agents on one account with no arbiter, where the first agent awake can drain the budget ^[10]. Of the three models it describes, only the centralized arbiter comes with code ^[11]. BudgetArbiter approves a request when the estimated cost fits inside the remaining balance, decrements the balance by that estimate, and later calls refund, which adds (estimated_cost - actual_cost) back ^[12]. Admission runs on the estimate and settlement runs on the actual. When the actual is larger, that expression is negative, the balance ends at remaining minus the actual cost, and it can go below zero ^[16]. The spread between the two numbers is whatever the price did in between ^[3].

The estimate is where the design gets thin. According to the post, the agent can call /v1/scry/estimate if Scry exposes one, and otherwise has to fall back on historical cost data or a conservative upper bound ^[13]. The documented primitive is /v1/scry/schema, which returns the live schema and the current price for each relation ^[4]; the cost of a query that already ran arrives in its response metadata ^[5]. A conservative upper bound reserves more than the query will cost and hands the difference back only after execution ^[18]. Until that refund lands, the reserved money is denying someone else's query ^[12]. The sketch guards the balance with an asyncio.Lock and keeps it in an instance attribute, so the ceiling holds for coroutines inside one process; ten agents in ten processes need the balance somewhere all of them can see it ^[17].

The price has to reach whichever component decides, because an agent cannot react to a signal the orchestrator does not pass down ^[14]. The agent has to carry a budget, or it will pay any price ^[14]. And the backoff has to key on the price, which the post prescribes; its worked example stops mid-sentence at "If the price is 10x the baseline" ^[9]. The post gives no price curve, no elasticity figure, and no measurement of the amplification it warns about ^[19].