Blog
Longer posts on how we think about RAG, agents, picking a model, and keeping AI running once it's in production.
Subscribe via RSS10 posts found.
Some systems will never get an API: the supplier portal, the benefits site, the government form. Claude in Chrome watches you do the task once and does it again itself, from inside your own logged-in browser.
Claude can now sit inside Excel, trace a broken formula back to the exact cell that caused it, and build next month's report from a plain sentence. Below is the workflow we use, plus the setting to check before you paste in real numbers.
CI goes red while you're in a meeting, and by the time you're back the failure has gone cold. Auto-fix watches the build for you and pushes a fix before you've even opened the tab.
A for loop calling the model once per ticket pays full price and competes with live traffic for the same rate limit. The Batch API processes up to 100,000 requests at once, at half the cost, for exactly the work that was never going to need an answer in three seconds.
Every new chat starts from zero. You paste in the same style guide, the same checklist, the same 'not like that' before Claude gets it right. Agent Skills let you write that knowledge down once, as a folder Claude opens only when the task actually needs it.
Every AI assistant your team adds needs the same connection to your order database, your CRM, your ticket queue, built again, in a different SDK. Model Context Protocol lets you build that connection once.
Retrieval-augmented generation lets a model answer from your own documents instead of guessing. How it works under the hood, and the cases where it's the wrong tool.
Chat is a conversation. Agents are a system that reads, decides, and acts, which means the engineering bar, and the failure modes, are completely different.
Leaderboards rank models on someone else's tests. In production you care about which one is cheap, fast and consistent on your task, and the answer is often a different model.
Getting a demo working is the easy 20%. This is the rest of the list: what a prototype needs before you'd let it run unattended in front of customers.