Scry: Congestion Pricing as Agent Rate-Limiting Infrastructure

Scry is a search API for agents that replaces hard rate limits with a congestion pricing model. Instead of receiving a 429 error when the quota is reached, agents receive a price signal. They can then decide whether to proceed, defer, or cancel the query based on the cost. This shifts rate-limiting from a strict binary limit to an economic feedback loop.

The system exposes an MCP server that allows agents to run SQL-like queries over 164 billion indexed internet documents, such as Reddit, Hacker News, and Wikipedia. Queries run with a deadline, memory limit, and row cap. Each query returns a cost in the response metadata, enabling agents to inspect the cost before execution if their SDK supports dry-run or estimate endpoints.

Unlike traditional rate limits, which are binary and result in a 429 error when exceeded, congestion pricing provides a gradient of prices. Agents can see the current price per query, historical price distribution, and remaining budget (if tracked by an orchestrator). They can execute queries immediately if the price is acceptable, defer them if the price is high and the task is not urgent, or cancel the query if the price exceeds the value of the answer.

There are trade-offs to consider. While congestion pricing offers excellent cost control and provides a price signal, it requires complex cost-aware decision logic from agents. If the orchestrator does not expose pricing signals, agents cannot react appropriately. Additionally, when multiple agents share a single billing account, congestion pricing creates coordination problems.

Three arbitration models exist: centralized arbiter, agent bidding, and no arbiter. The orchestrator must implement the chosen model, as Scry does not enforce one.

To prevent congestion amplification due to retry loops, agents should use cost-aware backoff instead of traditional exponential backoff tied to time. If a query fails due to high cost rather than a 429 error, the agent should back off exponentially based on the price, not time. This prevents many agents from hammering the API during price spikes and allows the system to dampen congestion instead of amplifying it.

Observability and cost attribution are also crucial when using congestion pricing. Agents need to track which queries are expensive, and orchestrators must be able to attribute costs to specific agents. This requires Scry to expose real-time pricing and query cost information.