DEV Community

Cover image for Anthropic is bullshitting you about subagents and loops.
Nabil A
Nabil A

Posted on

Anthropic is bullshitting you about subagents and loops.

Hear me, hear me. I had three Claude accounts, three at $200 a month each.

I was doing the whole thing: workflows, subagents, parallel development, loops, all the fancy bullshit. I'd burned through the available usage on all three accounts in about a week.

That's all three of them: $600 a month. Apparently I was sponsoring an AI group project without even knowing it.

When I look at the workflow I have a pretty good guess where a lot of that usage was going.

You've got agents working in separate Git workflows. Trees on separate branches, each making changes in its own little world.
Fine. but eventually all those changes have to work together in one codebase.

Now you're spending more tokens rereading code, reviewing the different branches, figuring out where changes overlap, resolving conflicts, and merging everything back together.

Guess what? If your agent resolves that overlap badly and someone's fix gets dropped or overwritten, you're screwed.

Even a clean Git merge doesn't mean the code automatically works. Two agents can make perfectly reasonable changes based on completely different assumptions about how the feature should behave. Congratulations.

Now you've got the opportunity and the privilege that Anthropic has given you to spend more tokens finding and fixing the problems created by putting everything together.

It feels like giving 10,000 engineers free rein to write code, then hiring another 10,000 reviewers to figure out what the first 10,000 did.

Bro, I'm trying to finish a feature.

So I decided to shift. I stopped developing everything in parallel. The planning phase can still happen in parallel: explore ideas, figure out the approach, whatever.

The actual development: one task, one thread. Finish it before moving on. You plan in phases and you develop sequentially.

Build it, test it, fix whatever broke, then start the next task.

That's it.

That's the methodology.

I know, I know. It's absolutely ground-breaking stuff but in my experience it actually works better: fewer errors, fewer bugs, more correct, more robust results. I'd estimate roughly 70% less usage on a single account.

Somehow it's actually faster too although after watching how much work goes into coordinating and cleaning up parallel changes, maybe that shouldn't surprise me.

Now, I know. I haven't run a controlled benchmark. These are observations from my own work and that 70% is really an estimate but I did delete two of my Claude accounts.

Now one account lasts me until the weekly reset. I'm getting my work done without managing a family phone plan for robots.

Also honestly, I feel like this is kind of intentional: Anthropic, the agent companies, the harness companies, pushing the same shit.

I feel like they're dressing up token consumption as productivity: more agents, more loops, more orchestration. You burn through your allowance and start thinking, "Hey guess I need another account." I fell into that trap twice.

Somewhere along the way we started treating the number of agents running as a measure of how much work was getting done. Meanwhile half the circus feels like it exists to manage the other half. I care whether the thing works, how long it takes, and how much usage I burn getting there.

Now some agents have their uses, but for most of my day-to-day development, the extra coordination wasn't really earning its keep.

Please, before you buy another account or pay for more usage, try sequential development (SeqDev) for a week: one task, one plan, build it in phases, test it, fix it, move on to the next phase.

See what happens to your usage, see what happens to your bugs, see how much you actually finish.

I tried it and fired two-thirds of my very expensive group project.

Top comments (0)