TOON: Cut AI Coding Agent Token Costs with Better Tool Output

AI coding agents spend tokens even before writing code, consuming them on source files, ticket descriptions, and other inputs. Most of the cost comes from what the agent reads, not the final pull request. Teams often focus on model choice and prompt length for cost control. However, a less visible factor is the format of data returned by developer tools to the AI model. Token costs are shaped by how development tools package information, not just what agents read. When a tool returns many similar records, verbose JSON can make the agent repeatedly process field names and structural syntax. For example, in an issue list, labels like "severity" and "component" are repeated for every record in conventional JSON. This repetition consumes context that could hold more relevant information. Token-Oriented Object Notation, or TOON, is one solution. It uses a schema-like header for uniform arrays, sending each record as a row. Field names appear once, while values remain. This approach provides the same findings with less repeated information. This matters because optimizing tool output formats can significantly reduce AI agent token consumption and costs.