Agentic Engineering – Tokens are not enough

Software Engineering is a wild place now. We’re doing more with less. We build sometimes twice as many features with leaner teams. PM’s write code, Engineers write code, design and do PM work.

Everyone is more productive and full stack in some sense of the word.

Where we’re at, a lot of us are stretching the limit of what we can do with harnesses like cmux, codex and claude. 1 Billion + tokens consumed a day is not unheard of. Hundred+ PR weeks.

In the context of this though – I want to discuss one important thing that I don’t see explicitly mentioned. Tokens are not enough. They’re a necessary, but insufficient condition in terms of Engineering output.

More than it has ever been – Engineering and Product in an organization exist to solve problems for customers and provide the backbone of a product that solves burning, real world needs. It’s important to be able to keep up with the rate of change in your organization – and for that you need to consume tokens – to monitor logs, to help with deeper reviews. To synthesize research. To stack PR’s and refine features quickly. You can’t do that with just your meat brain anymore.

But if you’re stacking tokens used and aren’t tracking the things that matter – SLOs (uptime), reliability, customer satisfaction. If you’re not building features that really matter – you’ll lose to teams that do. Build a lot – move fast. But you should build what matters.

It kind of reminds me of when I would watch competitive Starcraft 2 players. Some of the players were sooooo fast – reaching 300, 400+ APM. They’d dance with their SCVs and probes around the starting area. But then they’d move on and use those actions to build bases – to launch attacks and defensives maneuvers.

If you’re tracking a leaderboard and you’re not spending as much, or more time on customer outcomes, you’re like that player with 400 APM but you’re spending it ordering around workers in your starting base.