--- title: FinOps for AI Agents: Who Spent All the Tokens? — Tisha... description: A Microsoft-presented framework called Token Ops proposes moving AI cost management from request-level model gateways to run-level governance of agentic... canonical: https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft | daily.dev og:description: A Microsoft-presented framework called Token Ops proposes moving AI cost management from request-level model gateways to run-level governance of agentic... og:url: https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh og:image: https://api.daily.dev/og/posts/eC2Hma6fh.png og:image:alt: FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 21 min read · 1 upvotes · 0 comments ## Summary A Microsoft-presented framework called Token Ops proposes moving AI cost management from request-level model gateways to run-level governance of agentic workflows. The talk outlines first principles (token as unit of cost, attribution at the LLM call boundary, in-path steering before hard budget halts) and demos an out-of-band architecture with instrumentation, accounting, and enforcement layers. A live demo shows preview mode, hard-halt enforcement, and 'steer' behavior that injects instructions to curb token usage. Benchmarks against open-source agents (browser-use, MetaGPT) reportedly show a 78% reduction in average spend and a completion-rate uplift from 67% to 96% versus simple throttling. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much can run-level token governance reduce AI agent costs compared to simple throttling? Benchmarks on the open-source projects browser-use and MetaGPT showed average spend dropping by roughly 78% when a run-level governance system with steering policies (context compaction, tool output reduction, budget-aware prompt injection) was enabled, compared to no governance. Unlike simple throttling, which kills runs outright, this approach raised task completion rates from about 67% to 96% because it adjusts agent behavior before hitting a hard budget cap. _Teams weighing agent cost controls versus completion rates can track approaches like this on daily.dev._ ### Why don't existing LLM gateways like LiteLLM or Portkey fully solve AI agent cost overruns? Tools such as LiteLLM, Portkey, and Cloudflare enforce budgets at the individual model request layer, offering hard caps or model routing/downgrading. They lack visibility into the agent run as a whole, so they can't detect or control problems like runaway loops between an agent and its tools, uncontrolled spawning of sub-agents, or unbounded context growth, all of which happen above the single request boundary. _Developers debugging unpredictable agent bills can follow this kind of cost-control architecture on daily.dev._ ### What happened to Uber's AI budget that illustrates the risk of unmonitored agent token spend? Uber's AI budget reportedly got exhausted within four months, cited as an example of unbounded consumption in agentic AI systems. Other companies were said to have run up hundreds of millions of dollars in costs within months or even days due to runaway agent loops with no proper mechanisms in place to detect or halt the excessive spend before it accumulated. _Anyone budgeting for production AI agents can keep tabs on real-world cost incidents like this via daily.dev._ ## Similar posts on daily.dev - [Token Economics and TokenOps: The Definitive Guide to FinOps for Tokens](https://daily.dev/posts/token-economics-and-tokenops-the-definitive-guide-to-finops-for-tokens-piylswjrm) · finout · 1 upvotes · 0 comments - [AI FinOps: Why Token Economics Will Define the Next Generation of Enterprise AI](https://daily.dev/posts/ai-finops-why-token-economics-will-define-the-next-generation-of-enterprise-ai-tw7haawl1) · Security Boulevard · 1 upvotes · 0 comments - [AI Cost Visibility: How to Track and Optimize Token Spend Before the Invoice Arrives](https://daily.dev/posts/ai-cost-visibility-how-to-track-and-optimize-token-spend-before-the-invoice-arrives-133h7kulz) · Telerik · 1 upvotes · 1 comments - [Tokens Are the New Cloud Bill](https://daily.dev/posts/tokens-are-the-new-cloud-bill-cbzt2bq6s) · Cast AI · 0 upvotes · 0 comments - [FinOps for AI Tokens: Why the Rules Changed — and What to Do About It](https://daily.dev/posts/finops-for-ai-tokens-why-the-rules-changed-and-what-to-do-about-it-qg82j1ezl) · finout · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#observability](https://daily.dev/tags/observability), [#finops](https://daily.dev/tags/finops) [View this post on daily.dev](https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft","url":"https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh"},"datePublished":"2026-08-22T14:38:25.562Z","dateModified":"2026-09-13T20:35:03.274Z","description":"A Microsoft-presented framework called Token Ops proposes moving AI cost management from request-level model gateways to run-level governance of agentic...","image":"https://i.ytimg.com/vi/GJX19pNhmSw/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/GJX19pNhmSw/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,observability,finops","timeRequired":"PT21M","video":{"@type":"VideoObject","name":"FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft","description":"A Microsoft-presented framework called Token Ops proposes moving AI cost management from request-level model gateways to run-level governance of agentic...","thumbnailUrl":"https://i.ytimg.com/vi/GJX19pNhmSw/sddefault.jpg","uploadDate":"2026-08-22T14:38:25.562Z","duration":"PT21M","url":"https://api.daily.dev/r/eC2Hma6fh","embedUrl":"https://www.youtube.com/embed/GJX19pNhmSw"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/finops-for-ai-agents-who-spent-all-the-tokens-tisha-chawla-susheem-koul-microsoft-ec2hma6fh#faq","mainEntity":[{"@type":"Question","name":"How much can run-level token governance reduce AI agent costs compared to simple throttling?","acceptedAnswer":{"@type":"Answer","text":"Benchmarks on the open-source projects browser-use and MetaGPT showed average spend dropping by roughly 78% when a run-level governance system with steering policies (context compaction, tool output reduction, budget-aware prompt injection) was enabled, compared to no governance. Unlike simple throttling, which kills runs outright, this approach raised task completion rates from about 67% to 96% because it adjusts agent behavior before hitting a hard budget cap. Teams weighing agent cost controls versus completion rates can track approaches like this on daily.dev."}},{"@type":"Question","name":"Why don't existing LLM gateways like LiteLLM or Portkey fully solve AI agent cost overruns?","acceptedAnswer":{"@type":"Answer","text":"Tools such as LiteLLM, Portkey, and Cloudflare enforce budgets at the individual model request layer, offering hard caps or model routing/downgrading. They lack visibility into the agent run as a whole, so they can't detect or control problems like runaway loops between an agent and its tools, uncontrolled spawning of sub-agents, or unbounded context growth, all of which happen above the single request boundary. Developers debugging unpredictable agent bills can follow this kind of cost-control architecture on daily.dev."}},{"@type":"Question","name":"What happened to Uber's AI budget that illustrates the risk of unmonitored agent token spend?","acceptedAnswer":{"@type":"Answer","text":"Uber's AI budget reportedly got exhausted within four months, cited as an example of unbounded consumption in agentic AI systems. Other companies were said to have run up hundreds of millions of dollars in costs within months or even days due to runaway agent loops with no proper mechanisms in place to detect or halt the excessive spend before it accumulated. Anyone budgeting for production AI agents can keep tabs on real-world cost incidents like this via daily.dev."}}]} ```