--- title: How we make AI coding more cost efficient without... description: GitHub engineers explain why per-tool-call token reduction is the wrong efficiency metric for AI coding agents, since aggressive output shortening can trigger... canonical: https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How we make AI coding more cost efficient without sacrificing task quality | daily.dev og:description: GitHub engineers explain why per-tool-call token reduction is the wrong efficiency metric for AI coding agents, since aggressive output shortening can trigger... og:url: https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm og:image: https://api.daily.dev/og/posts/Ceg6jEPHM.png og:image:alt: How we make AI coding more cost efficient without sacrificing task quality 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. # How we make AI coding more cost efficient without sacrificing task quality **[GitHub Blog](https://daily.dev/sources/ghblog)** · 9 min read · 0 upvotes · 0 comments ## Summary GitHub engineers explain why per-tool-call token reduction is the wrong efficiency metric for AI coding agents, since aggressive output shortening can trigger costly recovery steps that increase total task cost. Four shipped changes in GitHub Copilot are detailed: a selective output compressor that preserves source-like output while trimming repetitive build/test/lint noise, removal of unused line-number prefixes in the file-view tool (cutting inference cost ~5% offline, ~3% online per user), a meta-prompting loop that halved the task-tool prompt while catching and fixing a regression that broke parallel sub-agent execution, and batched delivery of completed background task results to avoid extra retrieval turns (~2.3% AI Credit reduction). The overarching lesson: efficiency must be measured end-to-end across a full task and validated with online experiments, since a win in one workflow (like code review) can be a regression in another (like CLI). ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why did shortening tool output with something like RTK (Rust Token Killer) end up increasing total AI agent cost? Shortening individual tool responses can backfire because when omitted text mattered, the coding agent sometimes reopened the original output or reran the command to recover missing information. Those recovery steps added extra turns and carried more context forward, so the task used more tokens overall even though each individual response was shorter, in GitHub's benchmark evaluations with RTK. _Developers weighing token-shaving tricks for their own agents can track these efficiency findings on daily.dev._ ### How much did removing line-number prefixes from Copilot's file-view tool reduce AI inference costs? Removing line-number prefixes from the view tool's file output cut model-inference cost by roughly 5% in offline agentic coding benchmarks and reduced average daily model-inference cost per Copilot CLI user by about 3% in an online experiment. Line numbers had become unnecessary because current editing tools match surrounding code rather than targeting specific line numbers, so success rates and edit failures stayed unaffected. _Anyone tuning AI coding agent costs can follow real efficiency numbers like this one on daily.dev._ ### What went wrong when GitHub used a meta-prompting loop to shrink Copilot's task-tool prompt? A meta-prompting loop that let Copilot iteratively rewrite its own prompt cut it by roughly half, but the first online experiment revealed it had rewritten cautious parallelism guidance into a hard scheduling policy, causing independent custom agents to run sequentially instead of in parallel. The fix replaced an explicit allowlist and denylist with one sentence: independent agents can run in parallel, consider side effects. _Teams experimenting with self-optimizing prompts can track lessons like this on daily.dev before shipping._ ## Similar posts on daily.dev - [How we’re making GitHub Copilot smarter with fewer tools](https://daily.dev/posts/how-we-re-making-github-copilot-smarter-with-fewer-tools-euprmmzbg) · GitHub Blog · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#github](https://daily.dev/tags/github), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering) [View this post on daily.dev](https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm) ```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":"How we make AI coding more cost efficient without sacrificing task quality","url":"https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm"},"datePublished":"2026-09-02T18:01:04.775Z","dateModified":"2026-09-03T02:02:48.730Z","description":"GitHub engineers explain why per-tool-call token reduction is the wrong efficiency metric for AI coding agents, since aggressive output shortening can trigger...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6ba2f60e52e8020068ab76bbc8b720e3?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6ba2f60e52e8020068ab76bbc8b720e3?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"GitHub Blog","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":"GitHub Blog","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/106cf162b88840808484d4b5429b59b1","url":"https://daily.dev/sources/ghblog"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,github,ai-agents,prompt-engineering","timeRequired":"PT9M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"GitHub Blog","item":"https://daily.dev/sources/ghblog"},{"@type":"ListItem","position":3,"name":"How we make AI coding more cost efficient without sacrificing task quality"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-we-make-ai-coding-more-cost-efficient-without-sacrificing-task-quality-ceg6jephm#faq","mainEntity":[{"@type":"Question","name":"Why did shortening tool output with something like RTK (Rust Token Killer) end up increasing total AI agent cost?","acceptedAnswer":{"@type":"Answer","text":"Shortening individual tool responses can backfire because when omitted text mattered, the coding agent sometimes reopened the original output or reran the command to recover missing information. Those recovery steps added extra turns and carried more context forward, so the task used more tokens overall even though each individual response was shorter, in GitHub's benchmark evaluations with RTK. Developers weighing token-shaving tricks for their own agents can track these efficiency findings on daily.dev."}},{"@type":"Question","name":"How much did removing line-number prefixes from Copilot's file-view tool reduce AI inference costs?","acceptedAnswer":{"@type":"Answer","text":"Removing line-number prefixes from the view tool's file output cut model-inference cost by roughly 5% in offline agentic coding benchmarks and reduced average daily model-inference cost per Copilot CLI user by about 3% in an online experiment. Line numbers had become unnecessary because current editing tools match surrounding code rather than targeting specific line numbers, so success rates and edit failures stayed unaffected. Anyone tuning AI coding agent costs can follow real efficiency numbers like this one on daily.dev."}},{"@type":"Question","name":"What went wrong when GitHub used a meta-prompting loop to shrink Copilot's task-tool prompt?","acceptedAnswer":{"@type":"Answer","text":"A meta-prompting loop that let Copilot iteratively rewrite its own prompt cut it by roughly half, but the first online experiment revealed it had rewritten cautious parallelism guidance into a hard scheduling policy, causing independent custom agents to run sequentially instead of in parallel. The fix replaced an explicit allowlist and denylist with one sentence: independent agents can run in parallel, consider side effects. Teams experimenting with self-optimizing prompts can track lessons like this on daily.dev before shipping."}}]} ```