--- title: The Platform Engineering Playbook for Production LLMs description: Drawing on experience building an LLM-driven inventory recommendation system for a large retailer, the author argues that production AI agent reliability is a... canonical: https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The Platform Engineering Playbook for Production LLMs | daily.dev og:description: Drawing on experience building an LLM-driven inventory recommendation system for a large retailer, the author argues that production AI agent reliability is a... og:url: https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur og:image: https://api.daily.dev/og/posts/8cvqktVur.png og:image:alt: The Platform Engineering Playbook for Production LLMs 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. # The Platform Engineering Playbook for Production LLMs **[InfoQ](https://daily.dev/sources/infoq)** · 33 min read · 0 upvotes · 0 comments ## Summary Drawing on experience building an LLM-driven inventory recommendation system for a large retailer, the author argues that production AI agent reliability is a platform engineering problem, not an application-level one. Hallucination rates dropped from 15% to 1.5% not through a better model but by building shared infrastructure: structured output validation with Pydantic schemas, classified retry logic (schema vs hallucination vs infrastructure errors), a prompt registry with instant rollback, an intent-classification gate that returns 'unclassified' rather than guessing, default-deny authorization enforced at each MCP resource server, per-request token and hallucination attribution by team, and golden-set evaluations gating prompt promotion. The architecture uses Google's ADK for orchestration, LiteLLM for model-provider swapping, FastMCP for tool servers, and Ollama for a zero-cost local demo stack, with a full runnable reference implementation published on GitHub. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How can I reduce hallucination rates in a production multi-agent LLM system? Hallucination rate can be cut from 15% to 1.5% by enforcing structured output schemas (e.g., Pydantic models) at the platform boundary, classifying failures into schema violations, hallucination signals, and infrastructure errors, and routing each to a different retry strategy with a capped retry budget, rather than blindly rerunning failed calls or swapping the foundation model. _daily.dev surfaces engineering practices like this for teams hardening their own AI agent pipelines against hallucinations._ ### How do I prevent an intent-classification router from forcing a request to the wrong agent? Have the classifier return an explicit 'unclassified' status instead of defaulting to the highest-scoring agent when no intent clears a confidence bar. This prevents off-intent hallucinations because a weak match no longer gets routed to a specialist agent that wasn't built to handle it; instead the platform can ask the user to clarify or fall back to a constrained search. _developers designing multi-agent routing logic can find similar architecture patterns on daily.dev._ ### Where should tool authorization be enforced in a multi-agent MCP architecture? Authorization should be enforced directly at each MCP resource server with a default-deny policy, not only at the API gateway. Each MCP server verifies the caller's signed JWT and checks a YAML-based permissions policy before running any tool, so a single orchestrator bug cannot expose every connected tool across teams. _teams securing multi-tenant AI tool access can track RBAC patterns like this on daily.dev._ ## Similar posts on daily.dev - [Three Years of Building Agents in Production \(Part 2\)](https://daily.dev/posts/three-years-of-building-agents-in-production-part-2--9uvhtnda1) · Mabl Engineering Blog · 2 upvotes · 1 comments - [LLM hallucinations in production](https://daily.dev/posts/llm-hallucinations-in-production-bs191p5s1) · portkey · 1 upvotes · 0 comments - [Production-grade LLMs and agents: a field guide](https://daily.dev/posts/production-grade-llms-and-agents-a-field-guide-s3ajmurm3) · Stack Overflow Blog · 2 upvotes · 0 comments - [Stop failing on the path to production: A better way for agentic platforms](https://daily.dev/posts/stop-failing-on-the-path-to-production-a-better-way-for-agentic-platforms-kyp7odl86) · Temporal · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#observability](https://daily.dev/tags/observability), [#mcp](https://daily.dev/tags/mcp) [View this post on daily.dev](https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur) ```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":"The Platform Engineering Playbook for Production LLMs","url":"https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur"},"datePublished":"2026-10-05T11:01:35.394Z","dateModified":"2026-10-05T14:49:15.417Z","description":"Drawing on experience building an LLM-driven inventory recommendation system for a large retailer, the author argues that production AI agent reliability is a...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e78bbf7a3389066815d400e3a9948186?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e78bbf7a3389066815d400e3a9948186?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"InfoQ","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":"InfoQ","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/afc3bced3e1e4b188dd9127017a60e0c","url":"https://daily.dev/sources/infoq"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,observability,mcp","timeRequired":"PT33M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"InfoQ","item":"https://daily.dev/sources/infoq"},{"@type":"ListItem","position":3,"name":"The Platform Engineering Playbook for Production LLMs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-platform-engineering-playbook-for-production-llms-8cvqktvur#faq","mainEntity":[{"@type":"Question","name":"How can I reduce hallucination rates in a production multi-agent LLM system?","acceptedAnswer":{"@type":"Answer","text":"Hallucination rate can be cut from 15% to 1.5% by enforcing structured output schemas (e.g., Pydantic models) at the platform boundary, classifying failures into schema violations, hallucination signals, and infrastructure errors, and routing each to a different retry strategy with a capped retry budget, rather than blindly rerunning failed calls or swapping the foundation model. daily.dev surfaces engineering practices like this for teams hardening their own AI agent pipelines against hallucinations."}},{"@type":"Question","name":"How do I prevent an intent-classification router from forcing a request to the wrong agent?","acceptedAnswer":{"@type":"Answer","text":"Have the classifier return an explicit 'unclassified' status instead of defaulting to the highest-scoring agent when no intent clears a confidence bar. This prevents off-intent hallucinations because a weak match no longer gets routed to a specialist agent that wasn't built to handle it; instead the platform can ask the user to clarify or fall back to a constrained search. developers designing multi-agent routing logic can find similar architecture patterns on daily.dev."}},{"@type":"Question","name":"Where should tool authorization be enforced in a multi-agent MCP architecture?","acceptedAnswer":{"@type":"Answer","text":"Authorization should be enforced directly at each MCP resource server with a default-deny policy, not only at the API gateway. Each MCP server verifies the caller's signed JWT and checks a YAML-based permissions policy before running any tool, so a single orchestrator bug cannot expose every connected tool across teams. teams securing multi-tenant AI tool access can track RBAC patterns like this on daily.dev."}}]} ```