01How to Use Claude Code Agent Teams: The Ladder
If you want to know how to use Claude Code agent teams, start with what works today. Julian Goldie turned every rung of the agent team ladder that works right now into a Claude Code automation, then ran four of them for real on one business website.
Claude Code does not need loops anymore. You give it a goal and it keeps going until the job is done. So the next step is teams of agents.
The ladder has these rungs:
- One agent.
- Fan out.
- Effort scaling.
- Pairs.
- A manager thread.
- Long-lived leads.
- Swarms.
The four automations in this episode are:
- The bug bash swarm.
- Goal mode.
- Fan out research with effort scaling.
- Build, review and verify.
The test site was a demo dental clinic website. Six agents found 41 problems on it, and the other automations sped it up, compared tools for it and built a booking page for it.
02Automation 1: The Bug Bash Swarm
Julian starts at the top of the ladder with the one rung where a swarm already works: finding problems.
Six Claude Code agents start at the same moment. Each gets its own start page and its own lens:
- Links.
- SEO.
- Accessibility.
- Forms.
- Content.
- Speed.
What makes it a swarm, and not six separate agents, is one shared findings log. Before an agent reports a problem, the log checks whether someone already found it. If so, it marks it as a duplicate and the agent moves on.
Every few finds, each agent reads what the others found and follows the same lead on other pages.
After 46 seconds, the swarm had 41 unique problems, and eight duplicates were caught before anyone repeated the work. The finds included a missing booking page, 14 megabyte images, a script that locks up the page and contact form inputs with no labels.
Julian links this to Anthropic's research, where simple parallel agents found 21 security issues and a coordinating swarm found 266 using about four times the tokens. Swarms shine at finding things because every lead is independent.
03Automation 2: Goal Mode Beats Your Own Plan
Goal mode is the first rung, and Julian says most people still get it wrong. The usual way is to type a list of steps: speed up the homepage, compress the hero image, lazy load the cards, then move a slow script to the bottom.
That is a plan, not a goal. Here is what happened when the agent was given the outcome and a way to check it:
- Run 0: performance 61 and accessibility 73.
- Run 1: it compressed the images, the page dropped to 400 kilobytes, and performance rose to 72.
- Run 2 followed Julian's step three, moving the script to the bottom, and performance dropped to 53.
- The agent watched the check, not the plan, and found the real problem: a loop in the script that locked up the page for almost a second.
- It removed the loop and performance jumped to 93.
- It then fixed the fonts, the contrast, the phone layout and its own font files after one run came back flaky.
In total it ran 11 checks in about four and a half minutes and reached a performance score of 99.
The lesson: give Claude Code the outcome and a way to check, and it can beat your own plan.
04Automation 3: Fan Out Research With Effort Scaling
The third automation combines fan out and effort scaling. The demo clinic wanted more patient reviews and had five tools to compare: Google Business Profile, Trustpilot, Birdeye, Podium and NiceJob.
The prompt asks the fan out research skill to compare the five tools for a dental clinic on review requests by text and email, replying, showing reviews on the website, booking software links and setup effort.
Here is how the team splits the work:
- The lead runs on Opus and writes one brief.
- It starts five helpers at the same moment, one per tool.
- The helpers run on Haiku, the fast small model.
- Each helper reads the official sites, scores its tool from one to five and writes a source next to every score.
- Opus checks every score and builds a comparison board.
It is fan out because no helper needs another helper's answer. It is effort scaling because the reading goes to Haiku and the judging goes to Opus.
The run took 2 minutes and covered 34 pages and searches. Birdeye won for this clinic with 21 out of 25, because it was the only one that linked to the main dental practice systems, sent requests by text and email and had a website widget.
05Automation 4: Build, Review and Verify
The fourth automation is a pair: a builder and a checker. The sources from the research step tell the verifier what done means.
The flow goes like this:
- Opus builds the whole page in one go, with no tickets and no splitting it up.
- Codex reviews it.
- The builder fixes the real cause, because it already knows why it made every choice.
- Five verifiers start at once, one per flow, each running a real Playwright test in a browser.
- Three passed straight away and two failed.
- Both failures were mistakes in the tests, not the page, so the builder fixed the tests and reran them.
- All five passed.
It took about four and a half minutes from start to finish, and it works. In the demo, a patient books a time slot and gets a reference number, and when the next patient comes in, that slot is greyed out.
Julian ties this to Anthropic's harness research, where a builder and a checker pass work back and forth. As models improved, the checker mattered less on easy work but still helped at the edges. So spend verifiers on hard parts, like dates and double bookings, not on a contact form.
06Which Agent Team Rungs Work Today
Julian gives a clear scorecard for the ladder.
What works today:
- One agent with a goal.
- Fan out, when the pieces are truly independent.
- Effort scaling, with cheap models reading and strong models judging.
- Pairs, where the same builder should fix its own findings.
- A manager thread, one chat that starts and checks the others. Julian uses the Claude Code desktop app like this every day.
What does not work yet:
- Long-lived leads.
- Full swarms that build together.
He points to Anthropic's own research, which says agents work best when they treat other agents like tools, and they still struggle as long-lived peers.
So the setup that works right now is simple: swarms find problems and builders build. That is why the bug bash swarm only finds issues, and one builder fixes them. Pick the lowest rung that gets the job done, and only move up the ladder when the task really needs more agents working at once.
07How to Install and Run the Four Skills
Getting started is short. Copy the .claude folder into your project, then type the prompts you saw in the episode.
- Copy the .claude folder into your project.
- Type run the bug bash swarm to start the six finders.
- Type use goal mode and give it an outcome plus a way to check it.
- Type use the fan out research skill with your comparison request.
- Type use build, review, verify for a page that needs real tests.
One more tip from Anthropic's harness research: every part of a harness encodes an assumption about what the model cannot do on its own. So:
- Keep these skills short.
- When the next model ships, test them again.
- Delete the parts the new model no longer needs.
Common mistakes to avoid:
- Giving goal mode a step list instead of an outcome.
- Using a swarm to build, when swarms are only good at finding.
- Putting verifiers on easy flows instead of hard ones like dates and double bookings.
08Who Claude Code Agent Teams Are For
These automations suit anyone who runs or builds websites and wants Claude Code to do more than one job at once.
They are useful if you:
- Want a fast audit of a site for broken links, SEO, accessibility, forms, content and speed.
- Want to speed up a slow page without guessing the fix.
- Need to compare several tools with a source behind every score.
- Need a working feature, like bookings, that is tested in a real browser.
The key idea is to match the job to the rung. Use a swarm to find. Use goal mode to improve a score.
Use fan out when the pieces do not depend on each other. Use a builder and checker pair when getting it right matters.
Julian mentions that the bug bash swarm with its shared findings log, goal mode, the fan out researcher, the build, review and verify skill and the agent control room are inside the AI Profit Boardroom as ready Claude Code folders, with a short guide for each one and the faceless video skill that made the video.
