Published: June 07, 2026, 08:00 UTC
A new research paper quantifies a counterintuitive reality about AI software engineers: the most expensive part of having an agent write code is not the writing, but the reviewing. The study, titled “Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering,” finds that code review consumes nearly 60% of all tokens in agentic software development workflows.
What the study measured
Researchers at Stockholm University and Concordia University analyzed token consumption patterns in ChatDev, a popular multi-agent framework that assigns different AI agents to distinct software engineering roles. They ran 30 software development tasks using a GPT-5 reasoning model and mapped internal ChatDev phases to standard development stages: Design, Coding, Code Completion, Code Review, Testing, and Documentation.
The goal was straightforward: figure out where the money actually goes when you let AI agents build software autonomously.
The results
Code review consumed 59.4% of all tokens on average, dwarfing every other stage. Initial coding accounted for far less. The iterative back-and-forth of agents checking each other’s work, suggesting changes, and re-verifying drove the bulk of token consumption.
Input tokens made up 53.9% of total usage, meaning agents spend more time reading and processing context than generating new output. This suggests significant inefficiencies in how multi-agent systems share information, with each agent potentially re-processing the same context that another agent has already examined.
The paper describes these patterns as evidence of “potentially significant inefficiencies in agentic collaboration.” In plain terms, agents in the current generation of frameworks do a lot of redundant reading.
Why this matters for the industry
The findings arrive as the tech industry races to deploy agentic coding tools. TechCrunch recently reported on the broader scramble to manage AI’s runaway costs, noting that a new industry group is building a canonical framework for “tokenomics” including standards for AI token usage, billing, and new metrics like cost-per-intelligence.
The paper’s results suggest that companies deploying multi-agent coding systems may face unexpected cost structures. The expense is not in generating the first draft of code, but in the automated quality assurance loop that follows. Teams that optimize for agentic coding should focus on reducing redundant context passing and streamlining review protocols, not just improving code generation speed.
The methodology
The study mapped ChatDev’s internal phases to standard SDLC stages using a standardized evaluation framework. The researchers used GPT-5, a reasoning model, which produces additional “thinking” tokens as part of its output. The paper separately quantified input tokens, output tokens, and reasoning tokens across each development stage.
The paper was accepted at the Mining Software Repositories conference (MSR ’26), held in Rio de Janeiro in April 2026, and has received 3 citations since its January submission.
The takeaway
The “Tokenomics” paper provides empirical grounding for something many engineering leaders have suspected but could not prove: having agents collaborate like humans costs more than having a single agent do the work. The overhead of coordination, context sharing, and iterative review dominates the bill. Future research and tooling should focus on making agent collaboration protocols more token-efficient, or the economics of AI software engineering will remain worse than the demos suggest.
Sources: arXiv (January 20, 2026); TechCrunch (June 5, 2026); Salim, Latendresse, Khatoonabadi, Shihab (MSR ’26)