Matthew Berman published a video recently where he dropped the number: 2.54 billion tokens spent perfecting his OpenClaw setup. He laid out 21 use cases he runs daily. CRM ingestion from Gmail and Calendar. A knowledge base of articles, videos, and posts. A business advisory council of multiple expert agents. A security council that reviews code for vulnerabilities.
The list is a great inventory. But the interesting part isn’t the list. It’s what 2.54 billion tokens of running experiments forces you to confront.
#The decision underneath
When you’ve spent that much, you stop asking “can I make the agent do this?” and start asking “should the agent be doing this at all?”
Matthew’s 21 use cases are the survivors. They’re the things he kept after the experiments that didn’t earn their tokens got cut. Most of what gets tried in those billions of tokens never makes the list. That’s the actual lesson.
I run a much smaller setup. Hope on a Mac Mini, a few hundred million tokens to my name, mostly through her morning briefings and heartbeats. But I’m starting to feel the same pressure: the agent can do anything, so the question becomes which things are worth its time.
#My survivors so far
Things Hope does that have earned a permanent spot:
- Morning briefing at 7am - one Telegram message with weather, calendar, three priority items, and any overnight changes worth knowing
- Heartbeat triage every 5 minutes - cheap local Llama 3.2, only escalates if something crosses a threshold
- Memory maintenance - daily compression of raw notes into curated MEMORY.md
- Project context loading - when I open a workspace, Hope pre-loads the relevant CLAUDE.md and recent commits so I’m in flow faster
Things I tried and cut:
- Real-time mention monitoring across Twitter and YouTube. Too noisy, never actionable, burned tokens for nothing.
- Auto-drafting reply emails. Hope’s voice is good but not mine. The drafts always needed enough rewriting that I’d rather start from scratch.
- Calendar event creation from chat. Worked technically. Failed socially. I forgot what was on my own calendar because I’d outsourced putting it there.
The agent makes you faster at the things you should be doing. It also makes you faster at the things you shouldn’t. Knowing the difference is the actual skill.
#What Matthew’s 2.54 billion proves
You can’t shortcut this. The 21 use cases on his list look obvious in retrospect. They were not obvious before he spent the tokens.
What you can do is read his list, take the three that match your actual life, and try them. That’s not 2.54 billion. That’s maybe a hundred million. And you skip past most of the dead ends because someone already paid for them.
That’s the trade. Read his list. Run his survivors. Use the saved tokens to find your own.
Let’s go!