--- title: The Counting House | daily.dev description: Steve Yegge describes a pattern he calls the Counting House: using cheap, stateless AI models on tightly controlled rails to handle massive customer-service... canonical: https://daily.dev/posts/the-counting-house-hp7kiecje twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The Counting House | daily.dev og:description: Steve Yegge describes a pattern he calls the Counting House: using cheap, stateless AI models on tightly controlled rails to handle massive customer-service... og:url: https://daily.dev/posts/the-counting-house-hp7kiecje og:image: https://api.daily.dev/og/posts/hP7KIEcJE.png og:image:alt: The Counting House 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 Counting House **[Steve Yegge](https://daily.dev/sources/yegge)** · 43 min read · 2 upvotes · 0 comments ## Summary Steve Yegge describes a pattern he calls the Counting House: using cheap, stateless AI models on tightly controlled rails to handle massive customer-service workloads. Unlike his open, agentic Wheelhouse software factory, Counting House agents get no memory, no credentials, and can never write directly to systems of record. Work is scored intent by intent using a policy table, with five autonomy tiers (off, shadow, assist, contain, transact) and a confusion matrix to tie model mistakes to financial cost. A Saga Pattern orchestrator, often built on Temporal, handles all writes through validated, idempotent, multi-step transactions. He argues this pattern already works at banks even though most companies still struggle with AI inside software development itself. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What are the autonomy tiers in Steve Yegge's Counting House pattern for AI customer service? There are five tiers: off (humans handle it, no model involvement), shadow (model answers silently alongside humans for scoring only, no savings yet), assist (model suggests responses to a human agent before they act), contain (model answers the customer directly before a human sees it), and transact (model can make real changes like moving money, via an orchestrator). Each tier allows progressively riskier autonomy and is earned through evals recorded in a policy table. _Developers designing AI agent guardrails for production use can track emerging architecture patterns like this on daily.dev._ ### Why shouldn't AI agents write directly to a company's systems of record like databases or CRMs? Direct writes by models are error-prone, hard to undo, and leave a questionable audit trail, which matters for regulatory compliance requiring proof of who authorized a change. The recommended fix is having the model issue a 'write intent' naming the business outcome, which a mechanical orchestrator then translates into validated, idempotent, multi-step transactions using the Saga Pattern, commonly implemented with Temporal. _Engineers deciding how to safely wire AI agents into production systems follow orchestration patterns like this on daily.dev._ ### Why use precision and recall instead of accuracy to evaluate AI models on customer service tasks? Accuracy ignores the cost of different mistakes, so a 90.6% accurate model can still be unacceptable if its errors are expensive. Precision measures how many flagged cases were truly that intent, while recall measures how many true cases were caught; for money-related intents you want high precision, while for cases like special assistance requests you want recall near 100% so no one is missed. _Teams comparing model evaluation metrics for agent deployment can keep up with practical breakdowns like this on daily.dev._ ## Similar posts on daily.dev - [Trying the Software factory pattern.](https://daily.dev/posts/trying-the-software-factory-pattern--3wqdkev6u) · Irrational Exuberance · 3 upvotes · 0 comments - [How do you turn AI coding chaos into a repeatable playbook?](https://daily.dev/posts/how-do-you-turn-ai-coding-chaos-into-a-repeatable-playbook--pfj1d29zo) · Stack Overflow Blog · 2 upvotes · 0 comments - [The Agentic AI Handbook: Production-Ready Patterns](https://daily.dev/posts/the-agentic-ai-handbook-production-ready-patterns-zlhrkqtw7) · Hacker News · 1 upvotes · 0 comments - [The Multi-Agent Trap](https://daily.dev/posts/the-multi-agent-trap-gcmbn0u9m) · Towards Data Science · 1 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/the-counting-house-hp7kiecje) ```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 Counting House","url":"https://daily.dev/posts/the-counting-house-hp7kiecje","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-counting-house-hp7kiecje"},"datePublished":"2026-10-27T03:35:36.721Z","dateModified":"2026-10-27T03:36:14.347Z","description":"Steve Yegge describes a pattern he calls the Counting House: using cheap, stateless AI models on tightly controlled rails to handle massive customer-service...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/07d69d5c2a0bd10e375b8b690865ab6a?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/07d69d5c2a0bd10e375b8b690865ab6a?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Steve Yegge","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":"Steve Yegge","logo":"https://media.daily.dev/image/upload/s---P6HxL-S--/f_auto,q_auto/v1787723832/logos/yegge?_a=BAMAMicg0","url":"https://daily.dev/sources/yegge"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-counting-house-hp7kiecje","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,ai-agents,llm","timeRequired":"PT43M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Steve Yegge","item":"https://daily.dev/sources/yegge"},{"@type":"ListItem","position":3,"name":"The Counting House"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-counting-house-hp7kiecje#faq","mainEntity":[{"@type":"Question","name":"What are the autonomy tiers in Steve Yegge's Counting House pattern for AI customer service?","acceptedAnswer":{"@type":"Answer","text":"There are five tiers: off (humans handle it, no model involvement), shadow (model answers silently alongside humans for scoring only, no savings yet), assist (model suggests responses to a human agent before they act), contain (model answers the customer directly before a human sees it), and transact (model can make real changes like moving money, via an orchestrator). Each tier allows progressively riskier autonomy and is earned through evals recorded in a policy table. Developers designing AI agent guardrails for production use can track emerging architecture patterns like this on daily.dev."}},{"@type":"Question","name":"Why shouldn't AI agents write directly to a company's systems of record like databases or CRMs?","acceptedAnswer":{"@type":"Answer","text":"Direct writes by models are error-prone, hard to undo, and leave a questionable audit trail, which matters for regulatory compliance requiring proof of who authorized a change. The recommended fix is having the model issue a 'write intent' naming the business outcome, which a mechanical orchestrator then translates into validated, idempotent, multi-step transactions using the Saga Pattern, commonly implemented with Temporal. Engineers deciding how to safely wire AI agents into production systems follow orchestration patterns like this on daily.dev."}},{"@type":"Question","name":"Why use precision and recall instead of accuracy to evaluate AI models on customer service tasks?","acceptedAnswer":{"@type":"Answer","text":"Accuracy ignores the cost of different mistakes, so a 90.6% accurate model can still be unacceptable if its errors are expensive. Precision measures how many flagged cases were truly that intent, while recall measures how many true cases were caught; for money-related intents you want high precision, while for cases like special assistance requests you want recall near 100% so no one is missed. Teams comparing model evaluation metrics for agent deployment can keep up with practical breakdowns like this on daily.dev."}}]} ```