The AI advantage is shifting
The AI advantage moves from access to executionOpenAI published a number this week worth stopping on: 8.3×. That’s how much more output its most intensive enterprise customers now generate per active user than typical firms in their industry. The models matter, of course. But these companies aren’t pulling away because they found a secret model no one else can buy. They’re changing how work gets done. That’s the idea underneath Issue #001 of Worth Your Attention: The advantage in AI is shifting from access to execution. Here are three stories that make the case, plus one smaller piece of the emerging agent stack. 01 — Working with AIWhat happened According to OpenAI, “Frontier” firms now generate 8.3× more output per active user than typical firms, up from 2.6× in January. Agent use is also spreading beyond engineering: since February, weekly active Codex users grew 108× in legal, 41× in sales, 41× in recruiting, and 26× in marketing. This is real use happening in business units that have nothing to do with “technology.” Knowledge workers are finally getting real production value out of working with AI in the day-to-day. The direction is hard to miss: leading companies are asking AI to carry out more work across many different departments, not simply answer more questions. What changes Most companies can buy access to roughly the same intelligence. The separation happens in the change management that comes after the purchase. The companies pulling ahead are redesigning work around AI. They’re giving agents the context, tools, permissions, and review process to own an outcome instead of helping with one step. And while this is true, my mind goes to one key word: more. And more isn’t always better. You’re probably thinking, “Yay 🙄 So what, they’re using more tokens. Of course Codex would promote this!” We all know that token volume isn’t the same thing as valuable output. In this particular case “more” has two costs:
That raises the question: validation against what? What to do next “More” is a huge waste of time unless you get very clear on what “done well” in any given process looks like. Without that clear definition you might make eight versions of a thing, none of them good enough for you to share. Pick one thing you do at least weekly that eats a lot of your time. (You may already be using AI to draft or summarize or brainstorm portions of this but we’re going to take it a step further). First ask:
Note: Examples are better than instructions because humans are notoriously poor at describing all of the nuance of the work that we do. Collect them into a single place that your AI can access (Google Drive, SharePoint, or a local folder), so your AI can refer to these as examples of what that weekly thing looks like when it’s “done well.” Then open your agent of choice in AGENT mode (use Codex, Claude Code, etc.) and ask:
02 — Agents can use old applicationsWhat happened Computer use has been in public beta since October 22, 2024. Yesterday, Anthropic moved computer use, browser use, its Skills API, and its Files API into production. Anthropic cites an insurance-technology company whose longest claims workflow fell from 32 minutes to 13, with about 30% lower cost per task and 100% completion in its tests without changing their prompts. What changes A capable model is only the beginning. An agent also needs hands to operate the software (tools), instructions that capture how your team works (prompts), and access to the files required to finish the job (context). But those things don’t always live in places that are easily accessible to command-line agents. In some cases, we have legacy applications that are local or online that have no agent-facing interface at all. After all, they were designed years ago for humans to operate them. Computer and browser use allow the agent to use these legacy applications on your behalf and automate portions of the work that used to require human intervention. This change opens up many more possibilities to hand off the busy work of data entry to your agent teams. This is a bit of catch-up from Anthropic, which is trying to match the capabilities of computer use from OpenAI What to do next I almost always recommend building an agent interface (like MCP) into any software system you use on a regular basis so that agents can use it just like humans do. Review your interactions with legacy software and websites and ask:
Note: “safely” is doing a lot of heavy lifting. You still need a way to verify that this work was done well. See the previous section. Read the Anthropic announcement 03 — Intelligence becomes interchangeableWhat happened Stripe agreed to acquire OpenRouter, a platform that routes AI requests across 400+ models from more than 80 providers. OpenRouter can choose a model based on the task, price, speed, and reliability. Its founder put the point plainly: no single model will be best for every job. What changes A critical workflow shouldn’t depend on GPT, Claude, or Gemini alone. A single place that provides all of these with a fallback makes sure that your work keeps going. This announcement continues a trend toward the commoditization of frontier models and reinforces the importance of what you build around the model: your workflows, context, tools (like computer use), controls, and evaluation system. Models will keep changing. The harness is what you own and take with you. Increasingly, when we’re talking about the value that frontier model companies provide, we’re talking about applications like Codex and Claude Code, which make frontier models easy to use and deploy. Right now integrating your application of choice with an AI gateway like OpenRouter is still something best left to technologists. 🤓 What to do next Separate your workflow from the model underneath it. You want the option to route or swap models as price and performance change, without rebuilding the process that makes the output valuable. This may mean using a harness (software that manages an AI model) with an AI Gateway that provides access to all of the models from a single location. Or it may mean building a durable context system around a simpler file-and-folder structure that can be easily navigated by any model in any harness you choose. Ask yourself:
Also worth knowing — The agent stack is standardizingGoogle’s A2A protocol is moving into the Agentic AI Foundation alongside MCP. The simple version:
That gives the emerging stack a clearer shape: models power agents, MCP connects them to third-party tools, and A2A lets them coordinate with other agents. We’re all still very early. But open standards make modular, replaceable systems more plausible and reduce the amount of proprietary glue companies have to build. I don’t see A2A (Google) in regular use yet but I’m using MCP (Anthropic) all day every day. It’s nice to see us taking the best innovations from each of the major LLM providers. The first question used to be: Which AI should we buy? Start here instead: Which part of the work are we ready to redesign? What’s one task you’d like to hand off to AI? Hit reply and tell me. I read every reply. –KM |