I run dozens of Grok Bot agents. They handle my calendar, my inbox, product builds, bookkeeping, research and a lot of small chores I used to do myself. This month the token bill got my attention, so I treated each fix like a small bet: write down the change, write down what you expect, and check it a week later against real numbers.
I made about a dozen changes. Most of them I can't call yet. Three held up well enough that I'd recommend them to anyone running more than a handful of agents.
The first one is almost embarrassing. My bots shared an ops room where they posted status updates. A lot of those updates said nothing new. The problem is that in a group room, every post wakes every bot in it, and each wake costs a full model turn to read the room and decide there is nothing to do. So the rule became: if your status line says nothing changed, write it to a log file instead of the room. In the days before, the room averaged about 15 posts a day. In the week after, about 2.7. The status history is still there in the log files if anyone needs it. I didn't measure the token saving directly, because my usage export can't tell me which bot spent what. What I can tell you is that a lot of turns that produced nothing stopped happening.
The second was cadence. Bots love routines, and routines quietly multiply. One of mine checked an inbox three times a day and ran a separate related job on top of that. I merged them into one run every weekday morning. A weekday health check went to Monday, Wednesday and Friday. That removed about 26 scheduled runs a week. The part that matters is what didn't happen: no mail was missed in the inbox bot's log, and the health check didn't catch anything late. When I first planned the cuts I counted two more routines, but they turned out to be paused already, so they saved nothing. Check what's really running before you celebrate the cut.
The third was model choice for Cursor cloud agents, which do my repo work. The account default was a Fast mode that costs about twice the standard rate per token. Most launches never asked for anything else, so they inherited it. The rule now is simple: every launch passes the model explicitly with Fast off, and only a stated reason earns an upgrade. The day after the rule, the Fast share of my Cursor tokens went from 89% to 37%. By my estimate at list prices, that cheaper mix alone saved about 31% that day compared to running the same tokens on Fast. That is one day of data, so treat it as an early signal, not a final number.
Now the part I think matters most for founders. My Cursor token use dropped sharply over those days, and it would be easy to credit the model rule for all of it. It wasn't. Tokens per call stayed flat at around 146k. Most of the drop came from running far fewer agents, not from running cheaper ones. If I had only looked at the total, I'd have told you the wrong story. Split every saving into "did less" and "paid less per unit." They need different fixes, and only one of them is about which model you picked.
Even the rule that worked had a miss. One agent still launched on Fast and ran there for about three hours before a follow-up message moved it to the standard model. A rule that lives in one place doesn't reach every call that needs it. So now the model setting goes on every reply to a running agent too, not just the launch.
The rest of my changes are still open: lower effort settings, offloading research to short-lived workers, how bots hand each other work. Some look promising. I can't separate their effect from everything else that changed the same week, so I'm not going to claim them.
If you're starting your own agent setup, here is what I'd take from this. Kill the wakes that produce nothing before you optimize anything else. Count the routines that are really running, not the ones on your list. Set the model explicitly and never trust a default you didn't choose. And measure each change against a before number, a few days apart, with the confounders written next to the result. Cheap is mostly about not doing work nobody needed.
Give this to your Grok Bot
- If a bot's status would say nothing changed, write it to a log file instead of posting in a shared ops room where every post wakes every bot. Check: pointless "read the room, nothing to do" turns stop; status history still lives in the logs if you need it.
- Audit scheduled routines; merge overlapping jobs into fewer weekday runs; cut health checks to a lighter weekday cadence only where nothing time-sensitive is missed. Check: your inbox bot's log and health check still behave on time after the cut.
- On every Cursor cloud agent launch, pass the model explicitly with Fast mode off; upgrade only when you state a reason. Set the model on every reply to a running agent too, not just at launch. Check: your mix shifts away from Fast; when totals drop, separate "did less work" from "paid less per unit" before you credit the model rule.