AI coding agents consume far more tokens—and therefore cost—on reading information than on generating code, according to a technical analysis published by The New Stack. Before an agent produces a single line of code, it's already spent tokens on source files, ticket descriptions, build logs, quality findings, and dependency alerts. Most of what teams pay for isn't the pull request itself, but everything the agent had to read to get there.
The cost problem centers on how developer tools package information for agents. When tools return large lists of similarly structured records, verbose JSON forces the agent to pay repeatedly for field names, quotation marks, and structural syntax with every entry. In a real-world comparison using 25 issues, Token-Oriented Object Notation (TOON)—a format that lists field names once in a header and then streams values as rows—used 49% fewer characters than pretty-printed JSON and 33% fewer than minified JSON. An agent reviewing 25, 100, or 500 findings doesn't need to be told the meaning of "severity" or "component" hundreds of times, yet conventional JSON repeats those labels with each record. That repetition eats context that could instead hold more relevant evidence, instructions, or source code.
The report finds that token costs are determined not only by what coding agents read, but also by how development tools package that information. For agentic workflows that involve agents calling tools in loops—listing findings, inspecting files, making changes, running analysis, and listing remaining findings—a modest reduction in one response compounds across repositories and iterations. The analysis cautions that "a cheaper context that causes a weaker decision is not a cost improvement," emphasizing that any format must preserve the fields the agent needs to make sound decisions and be validated against tasks like identifying the highest-priority finding and determining whether remediation is complete.
The report explains that teams typically focus cost controls on model choice, prompt length, and request limits, but overlook a less visible lever: the format of data returned to the model at the interfaces between agents and developer tools. The output format is an engineering decision, not a cosmetic one, because repeating structural overhead in uniform collections wastes context without adding information. A regular structure gives a model an explicit set of fields to expect, which can make review and validation more predictable. The report recommends using a table when a person needs to scan a short result in a terminal, standard JSON when a script or deeply nested payload benefits from its familiar structure, and a compact schema-first representation when an LLM is reading many records with the same fields.
The report advises teams to start with the highest-volume structured call in an agent workflow, measure the baseline, change one format setting, and then assess token consumption, response quality, and task completion together. That approach avoids a platform rewrite and makes the trade-off visible. As AI-assisted development becomes routine in engineering work, disciplined choices about what agents see and how they see it will be as important as the models themselves. The narrow, practical question isn't whether to abandon JSON entirely, but whether tools can return the same information in a representation designed for the shape a model needs to consume. Organizations that treat tool responses as part of agent design—not just model configuration—stand to reduce unnecessary context without compromising the quality bar for code or security findings.