Skip to main content
Markets · Live
DJIA50,687.07▼ -1.21%S&P 5007,553.68▼ -0.74%Nasdaq26,853.98▼ -0.89%Nvidia—▼ TSMC—▼ Alphabet—▼ Bitcoin—▼ DJIA50,687.07▼ -1.21%S&P 5007,553.68▼ -0.74%Nasdaq26,853.98▼ -0.89%Nvidia—▼ TSMC—▼ Alphabet—▼ Bitcoin—▼
New York--:--
Chicago--:--
Denver--:--
Los Angeles--:--
Honolulu--:--
Anchorage--:--
Tokyo--:--
Berlin--:--
London--:--
UTC--:--

Breaking

MIT TECH REVIEW — Stop counting tokens, start counting what ships · Nov 22, 8:30 AM ET
MIT Tech Review

Stop counting tokens, start counting what ships

When a developer gives an AI agent a coding task, the agent does more than answer once. It plans, generates code, runs tests, and repeats the process to correct errors. A single task can use 10 to 50 times more tokens than a basic chat prompt. The more useful the agent becomes, the more calls…

Affan Syed
By Affan Syed· MIT Technology Review· Published

Rights: fair use excerpt

No AI assistance disclosed

Stop counting tokens, start counting what ships
Reporting by Affan SyedSource: MIT Technology ReviewUpdated November 22, 2026
PHOTO: MIT TECHNOLOGY REVIEW

Executive Summary

Synthesized by V&V editors

The metrics governing enterprise artificial intelligence are undergoing a fundamental reassessment as autonomous coding agents replace static conversational tools in software engineering workflows. Writing for MIT Technology Review, technologist Affan Syed argues that organizations must shift away from tracking raw token consumption and instead evaluate machine learning investments by the actual software delivered. This conceptual transition reflects the evolving architecture of developer tooling, where computational efficiency matters far less than functional, finished output.

Traditional cost models established for standard conversational prompts fail to capture how autonomous software agents operate, according to the outlet. Rather than delivering a single transactional answer to an engineer's initial prompt, contemporary agentic architectures execute complex, multi-stage assignments independently. Upon receiving an objective from a programmer, an autonomous system systematically outlines an implementation plan, generates the necessary source code, executes automated testing suites, and iteratively alters the material to resolve identified errors.

Because these programmatic workflows rely on repetitive cycles of evaluation and debugging, their demand on underlying language models expands dramatically. MIT Technology Review notes that carrying out a single engineering task via an autonomous agent can consume between ten and fifty times the volume of tokens required by a standard interactive chat query. As an agent becomes increasingly adept at navigating intricate problems, it naturally initiates a higher frequency of model calls, accelerating token usage in direct correlation with its utility.

This compounding computational activity presents a tactical dilemma for leadership teams seeking to balance infrastructure outlays against developer productivity. Conventional procurement approaches that classify processing tokens solely as an operational cost overlook the substantive business value produced when autonomous agents eliminate friction from the development pipeline. The analysis indicates that tracking inputs becomes counterproductive if the intensive compute cycles successfully resolve technical debt and advance project delivery.

For executives, enterprise founders, and operational leaders, the emergence of iterative AI agents underscores the necessity of aligning performance indicators with core mission milestones. Managing high-capability tools requires shifting strategic oversight away from restricting processing capacity and toward measuring whether automated systems reliably deploy working code. Long-term operational superiority will belong to organizations that measure technology investments not by raw usage fees, but by their capacity to accelerate execution and deliver mission-ready solutions.

This Executive Summary is an original synthesis by Valor & Ventures Media editors based on public reporting by MIT Technology Review. For the complete original article, please visit the source.

---

Affan Syed · MIT Technology Review

Source & Credit

Reporting and photography credited as noted above. Originally published by MIT Technology Review.

Members only

Checking your membership…