What 100B+ Claude tokens actually look like inside a tiny company
According to my own usage tracking, I have used more than 100 billion Claude tokens since January.
That number is Claude only. It does not include Codex, Gemini, or the agent workflows we run through our server APIs. The same tracking showed 16.9 billion Claude tokens in the last 30 days.
The number is less interesting than what it represents.
We are a very small team, and we no longer operate with a traditional department structure. Instead of having separate teams coordinate every task through meetings, we have been building an automated system around code, marketing, and distribution.
The system tries to keep working around the clock. "Tries to" is doing a lot of work in that sentence.
In practice, the human role has shifted toward:
- Setting direction and defining the outcome
- Supplying context
- Reviewing what the agents produce
- Correcting course when they take a wrong turn
- Deciding what is worth doing next
It has not removed the need for people. It has changed where our time goes.
The trade-off is real. It is harder to switch off when work can continue without you. Sometimes there is useful progress waiting in the morning. Sometimes there is a confidently wrong implementation that needs to be unwound.
The upside is flexibility. I can spend uninterrupted time with my family and return for focused check-ins instead of manually carrying every task.
That is what an agent-first setup feels like for us right now. It is not a claim that agents can run a company without humans. It is a practical attempt to move human effort toward judgment, context, and strategic decisions.
For those already using agents in production: what have you delegated, and where do you still keep a human in the loop?
The attached screenshot shows my Claude usage over the last 30 days:
- 16.9B tokens across 237 sessions and 63,996 model turns
- File reading and editing as the largest category
- Heavy browser automation and terminal command usage
- Relatively little direct messaging compared with the tool and context work
For this account, most of the usage is context gathering and execution, with relatively little manual input. That distribution is the practical shape of the agentic setup I am describing.
We are building and advertising multiple products at once with this type of workload, our main one is https://brightbean.xyz/
Really impressive setup, Jan. Congrats on building this with such a small team, and thanks for sharing the real trade-offs instead of just the upside.
I think the biggest shift here is that the bottleneck moves from execution to judgment. Research, drafting, data movement, and repetitive tasks can be heavily delegated, but human checkpoints still matter around irreversible actions and customer-facing decisions.
At this scale, designing the right review system probably becomes just as important as the agents themselves.
Most forget when you yourself build you do a thousand checks. Why should it be different with AI? Plan, build, check, recheck, test, rinse and repeat.
I see reading and editing is your main token usage. I build a little tool for my repo's that makes a database with all references so AI can find things quick without going through thousands of lines. It cut my token usage for reading significantly. Might be usefull.
Mostly delegated the drafting grind - first-pass research, structuring, turning it into an actual finished deliverable. Kept myself in the loop on anything that goes out publicly or touches a live account: final review before publishing, pricing/positioning calls, that kind of thing. The "confidently wrong" risk you mention is real, especially the first few times you let it run further than you're used to - worth checking its output against reality before it snowballs.
The shift from doing the work to reviewing and correcting agent output seems like the more interesting change here.
I’m curious where you’ve found the boundary is hardest to define — which decisions can safely run without you, and which still require human judgment?
If it comes to development work, having a coherent user flow through the applications that are being build is still very tricky for the agents to figure out correctly. Working on it to have specific agents for this as well.
That makes sense. The user-flow problem feels like a different class of challenge from individual coding tasks, especially when agents can produce technically valid pieces that don't add up to a coherent experience.
I’d be interested in continuing the conversation outside the thread. What’s the best email to reach you on?
The shift from doing to reviewing and correcting is the real change. Most people underestimate how much judgment is still needed — agents multiply your output but they don't replace the decisions about what's worth building.