Controlled chaos: the pro mindset for week three
Welcome to the finale week. We are stepping off the well-lit path of one-you-and-one-agent and into the wilder territory of many agents at once. It is exciting, a little chaotic — and it works best when you understand one balancing act: controlled chaos.
Last week you focused on stage five of the eight stages — one agent, one you, letting diffs scroll by while you kept a watchful eye. This week we push into stages six, seven and eight: multi-agents, then many agents (swarms), then orchestrating agents. Bigger power, bigger stakes.
01A week to remember
Here is the shape of the finale: today, the pro toolkit — slash commands, sub-agents, hooks and plugins. Then sandboxing (running the craziness safely). Then working across large codebases. And finally the wild day: swarms, orchestrators, and a capstone project built almost entirely by agents.
Everything this week falls into two balancing camps. Chaos techniques amplify what’s happening — more agents, more automation, more surface area. Control techniques rein it in — shared files, self-correction, sandboxes, orchestration. Master coding is knowing how to balance the two.
02Two opposing forces
Think of it as positive feedback (things get bigger and bigger) versus negative feedback (things get pulled back on track). You want both, in tension.
Concretely: using the file system as shared memory so every process reads and writes the same architecture notes, open questions and test strategies; giving some agents the job of reviewing and testing others; running everything inside a sandbox you can throw away; and adding orchestration so one agent coordinates the rest. These are the brakes.
03The nuclear-reactor rule
The mental image for the week: a nuclear reactor. A chain reaction is enormous energy — but you lower control rods to keep it from running away. That is exactly the posture for pro coding agents.
Controlled chaos. Amplify what your agents can do (the chain reaction), while keeping enough control rods in place (reviews, files, sandboxes, orchestration) that it never melts down.
04Start with control, then loosen
How much chaos versus control? It depends on how mature your project is and your appetite for risk. The reliable recipe:
Begin with controldefault
On a real project, start tight — shared plan, close review. (Feel free to run a throwaway swarm or Ralph loop first just to explore the shape.)
Loosen graduallyas trust grows
When results look good, allow more autonomy — more agents, more automation.
Re-tighten if it driftssafety valve
If things start going off the rails, add control back. The dial moves both ways.
No matter how many agents you unleash, the code is your work product. The agents accelerate you; they do not absolve you. Keep a hand on the control rods.
05Today’s toolkit
The rest of Class 12 equips your belt with five pro features. You will rarely need all of them at once — but knowing they exist, and roughly how they work, is the point.
Custom slash commands
Your own /shortcuts for repeated prompts.
Agents & sub-agents
Delegate a task to a separate agent with its own context — even a different LLM.
Hooks
Make Claude Code do something automatically when an event fires.
Plugins & marketplaces
Bundle it all and share it with your team.
Before the tools, we set up the week’s capstone: FiNALLY, a multi-agent trading workstation. Every technique this week gets tried out on it. That’s next.
✓ Key takeaways
- Week 3 climbs from one agent (stage 5) to multi-agents, swarms and orchestration (stages 6–8).
- Pro techniques split into chaos (amplify) and control (rein in) — you balance them.
- The image is a reactor: huge energy, kept safe by control rods. Aim for controlled chaos.
- Recipe: start with control, loosen as trust grows, re-tighten if it drifts.
- Today’s belt: slash commands, sub-agents, hooks, plugins — tried out on the FiNALLY project.