First thing after vacation, I brought 100 bucks to Anthropic to play with their agent with fresh energy. For those who don't know - they have Max subscriptions for $100 and $200, which include the standard $20 for apps plus 5x/20x limits on Claude Code. And this is essentially the only tangible advantage over the approach with a bunch of MCP on top of Claude Desktop, which Vanya was talking about yesterday at the NPP meeting.
I'm almost sure that at my pace, it would be almost impossible to just spend this money on tokens. My pace is due to the fact that at work I'm either a solver in terms of current tasks or a brain in terms of some future. I don't write a lot of code as a typical product developer. But I wanted to try how I behave without any limitations at all. After all, psychologically these are two big differences - when you pay for each next word and when you've already paid for a subscription to this thing in advance. In the first case, you subconsciously save and don't go ask the agent a question where you could have asked it. In the second, you're constantly spurred by the feeling that you're underusing what you paid for.
And the result surprised me in principle. Because the most useful use case turned out to be not the most obvious before the experiment started.
This devil machine fantastically answers questions about the project or about some technical bottlenecks. It helps me quickly navigate tasks where I'm not in context at all. And I'm almost never in local context. Like "take classes A and B, see why this test doesn't pass". Or "they're asking me what fields we send in request C, prepare an answer". Or "tell me how in large Android projects they usually do D, give recommendations in the context of the current project". Or "figure out why when we publish release notes the message is formed not as needed, it should be in format E".
And in general, this applies to any task where I don't need a formal result (read - code). Roughly in this same area is writing documentation. If I need to record some thought, I'm almost completely satisfied with a draft from Claude Code. Like "let's write documentation on how we write unit tests". It goes to look at the tech stack itself, finds existing tests itself, extracts everything common from them, writes documentation by analogy with everything else in the project, and the result is really good. And if you also supplement the prompt with a guide on how to write this documentation, ask it to ask clarifying questions - it's a bomb.
That is, this is all some kind of service work, not its essence in the sense of writing code. I mainly write code with it as part of some experiments only. Where I wouldn't have had time to write code in my schedule in principle. And the quality of this work still doesn't amaze me. Partially with prompts/documentation for the task, but not easily.
And now it seems to me that this is exactly how everyone should use agents first of all. Literally, from intern to leads, what unites us is that we can get knowledge faster and in the context of a specific project. Not a ready result of work, but knowledge that secondarily converts into work results and just into your own development.
In the comments I'll add a screenshot with a table of tokens spent over these two weeks, also quite an interesting sight. It's definitely incredibly profitable compared to API, but the psychological barrier of $100 is hard to overcome.