Agentmaxx
TL;DR
TL;DR: I built and open-sourced agentmaxx, a Claude Code skill that will make long scraping, maths, or complex business problems take 1/10th of the time.
It has Claude split big jobs across lots of agents running at the same time, instead of working through them one by one. Before each run it checks how much your laptop can take. You can install it just by pasting the GitHub link into your Claude Code / IDE.
Intro
I was adding sources to Clara Jobs, the job feed I built for my wife. The plan was about 100 European universities, with each one having its own scraper. Three hours in, the progress was minimal so I asked Claude what it was doing, turns out it had only one agent open, doing one uni at a time. I then asked it why it wasn't running 50 agents at once, which to me seemed the logical way of doing it.
Why Claude only used one
It gave me three reasons as to why it didn’t default to a multi-agent workflow.
Claude Code only starts a multi-agent workflow if you ask for one in the prompt, by typing "ultracode" or asking it to spin up agents. Even though I’d mentioned I prefer parallel agents, unless you explicitly state what you want for this task, it will run singular ones.
Even when you do ask, the default size is under 10 agents. You can raise that in /config under "Dynamic workflow size", which I didn't know existed.
And a workflow runs at most 16 agents at once, or your number of CPU cores minus two if that's lower. My Mac has 15 cores, so that's 13 at a time, and the rest wait in a queue. I'd assumed all of this happened in the cloud. The model does, but each agent is a process on your laptop, and its commands, scrapers and builds run locally.
What slows a laptop down
The agents themselves are cheap. When I checked, 11 Claude processes were using about 1.75 GB of memory between them, around 150 MB each, so fifty would take 7–8 GB of my 24. The problem is the tools they run. A headless browser, which a scraper uses to load a page the way a person would, takes 300 MB to 1 GB each, and twenty of them would fill my memory. Builds, npm install and test suites each use a whole core while they run.
Either way, the dumb cap on the number of agents that Claude puts in place by default is silly.
What agentmaxx does
After prompting /agentmaxx every task runs as a multi-agent workflow without asking you each time. Small edits and quick replies stay with one agent.
Before each run it checks the machine's cores, memory and current load, and works the caps out from them: memory in GB divided by 4 for browsers (at most 6), and cores divided by 4 for builds. On my Mac that's 6 and 4. If the laptop is already busy, it halves both.
Before it launches, it tells me in one line how many agents it's starting, which heavy tools they'll use, and in what batch sizes.
The first run
Be aware, this eats usage. On my first run all 110 agents ran on Opus, the most expensive one. Fifty-five minutes in I hit my 5-hour usage limit, and 92 of the 110 stopped with "You've hit your session limit".
So the skill now sets a model for each agent: Sonnet for building and checking, Haiku for small steps like loading files, and Opus only in the main session, where the decisions get made. To get past 13 at a time, it splits the work into slices and runs several workflows side by side. The rerun split the 91 universities into three slices and ran 39 agents at once. It still hit the usage limit once, and carried on after the reset. By the end of the next day Clara Jobs had 84 working university scrapers and 113 sources covering 180 institutions.
Where it is now
I find myself using /agentmaxx as the default for any intensive tasks, it’s probably the most useful thing I’ve built for the Claude workflow. It's saved me hours.
It's on GitHub as agentmaxx. It works on macOS and Linux, you can install via curl, or as I simply do with all GitHub repos, simply paste the repo into your Claude Chat.