{"slug": "firebase-says-its-agent-skills-nearly-tripled-ai-coding-success", "title": "Firebase Says Its Agent Skills Nearly Tripled AI Coding Success", "summary": "Firebase's Agent Skills library raised a coding agent's task pass rate from 31.7% to 78.0% in the company's own evaluations, according to an Aug 11 blog post by product manager Charlotte Liang. The gain came with lower token counts and faster execution, and Firebase recommends pairing Agent Skills with its MCP server for best results.", "body_md": "# Firebase Says Its Agent Skills Nearly Tripled AI Coding Success\n\nFirebase published the eval numbers behind its Agent Skills library: pass rate jumped from 31.7% to 78.0%, with lower token and time costs.\n\nFirebase published the evaluation methodology behind its Agent Skills library, and the headline number is stark: giving a coding agent Firebase’s skills raised its task pass rate from 31.7% to 78.0%, according to a [firebase.blog post by product manager Charlotte Liang](https://firebase.blog/posts/2026/08/eval-driven-development-agent-skills), published Aug 11.\n\n## What Firebase measured\n\nAgent Skills, which Firebase [shipped in February](https://firebase.blog/posts/2026/02/ai-agent-skills-for-firebase/), are structured instruction files that tell coding agents like Claude Code, Cursor and Gemini CLI how to use Firebase correctly: which APIs are current, which are deprecated, and how to avoid common mistakes. The August post explains how the team decided which instructions to write and ship: build the eval first, run the agent without skills to capture a baseline failure rate, then add and refine skills to close the specific gaps the baseline exposed.\n\nThat baseline-to-skilled gain, 31.7% to 78.0% pass rate, came with what the post describes as substantial reductions in token overhead and execution time. Firebase breaks tokens into two buckets it tracks separately: input tokens (the prompt, the skills, tool definitions, error messages, web searches) and output tokens (the agent’s code and reasoning). Fewer input tokens signals the agent spent less time hunting for context; fewer output tokens signals a more direct answer with less rewritten code.\n\nThe post also draws a line between Firebase’s two agent-facing tools: the CLI, which agents favor for its straightforward terminal execution, and the MCP server, whose explicit tool definitions make it easier for an agent to know exactly what a given call does and expects. Firebase’s framing is that Skills are the layer that coordinates both, and that the best eval results came from enabling Agent Skills and the MCP server together rather than either alone.\n\n## Why it matters\n\nEvery AI-tooling vendor claims agents work better with its product. What’s different here is that Firebase published the eval discipline behind the claim, not just the claim: write the eval before the skill, treat a stuck agent as a signal that the underlying CLI or MCP command is too fragile rather than that the skill needs more words, and keep iterating the tooling itself as failures surface. A team weighing whether to write its own agent instructions can reuse that process directly; the top-line percentage alone gives them nothing to copy.\n\n**Firerun’s take:** a 31.7%-to-78.0% jump is Firebase’s own benchmark, on Firebase’s own tasks, and should be read as a best case rather than a number that transfers to another platform’s Skills. What is transferable is the admission buried in the methodology: a skills file nobody measures is close to guesswork.\n\n## Key Takeaways\n\n- Adding Firebase’s\n**Agent Skills library** to a coding agent raised task pass rate from**31.7% to 78.0%** in Firebase’s own evals, per the Aug 11[firebase.blog post](https://firebase.blog/posts/2026/08/eval-driven-development-agent-skills). - The gain came with\n**lower input and output token counts and faster execution**, tracked as separate metrics for context-search efficiency and answer conciseness. - Firebase’s recommended setup pairs\n**Agent Skills with its MCP server**, rather than shipping instructions without the tool definitions an agent needs to act on them. - The methodology, write the eval first, baseline without skills, then iterate, is offered as a template other teams can reuse beyond Firebase.", "url": "https://wpnews.pro/news/firebase-says-its-agent-skills-nearly-tripled-ai-coding-success", "canonical_source": "https://firerun.io/firebase-eval-driven-agent-skills-2026/", "published_at": "2026-08-19 00:00:00+00:00", "updated_at": "2026-08-19 10:11:00.381986+00:00", "lang": "en", "topics": ["ai-tools", "ai-agents", "developer-tools", "ai-research"], "entities": ["Firebase", "Charlotte Liang", "Agent Skills", "Claude Code", "Cursor", "Gemini CLI", "MCP server"], "alternates": {"html": "https://wpnews.pro/news/firebase-says-its-agent-skills-nearly-tripled-ai-coding-success", "markdown": "https://wpnews.pro/news/firebase-says-its-agent-skills-nearly-tripled-ai-coding-success.md", "text": "https://wpnews.pro/news/firebase-says-its-agent-skills-nearly-tripled-ai-coding-success.txt", "jsonld": "https://wpnews.pro/news/firebase-says-its-agent-skills-nearly-tripled-ai-coding-success.jsonld"}}