Anthropic just confirmed why 90% of non-coding AI agents fail in production

` tags and tracking scripts until it hallucinates. The developers getting agents to work in those low-percentage brackets on Anthropic's chart (like automated market research or live CRM routing) are usually spending most of their time on the boring infra work behind the scenes such as clean inputs, reliable scraping and that’s the part that really makes the difference. If you look at a modern, high-reliability agent stack outside of coding, it usually relies on three things: 1. The Core Reasoner: Something fast with a massive context window like Claude Sonnet to handle the logic. 2. Data Hygiene at the Gateway: Instead of letting the agent scrape raw web URLs directly (which triggers bot blocks and inputs HTML that will need to be revised), developers feed the internet data through dedicated markdown converters with tools like Firecrawl or Jina Reader are pretty standard here and the agent gets pure text, saving token costs and preventing hallucinations. 3. The Guardrail Layer: Traditional code hooks or rules engines that check the agent’s output before it executes an irreversible action (like sending an email or updating a database record). The low adoption numbers in the rest of the enterprise doesn’t mean agents are overhyped. In most industries, the surrounding tooling just still kind of sucks so once the data side gets more reliable, you’ll probably see adoption spread a lot faster outside engineering What are your thoughts on this? For those building agents in finance, marketing, or operations, I would love to get your thoughts here!