Multi-agent systems · Claude Code subagents
Open-source "AI agencies" ship a roster of specialist subagents — engineering, QA, design, marketing, ops — that drop straight into a coding tool like Claude Code. Under the hype is a real architecture pattern: specialists with isolated context, coordinated by one orchestrator. Grow the org below and watch the numbers.
Drag to rotate · top node = orchestrator
Isolation on: each specialist sees only its own brief. Turn it off to watch shared context balloon.
A subagent is basically a markdown file: a name, a scoped role ("you are the database migration reviewer"), allowed tools, and conventions. Narrow instructions measurably beat one 5,000-word do-everything prompt, because the model isn't juggling sixteen jobs at once.
Each specialist runs in its own context window and returns only a summary. The orchestrator never drowns in raw file dumps, and one agent's confusion can't contaminate another's reasoning — the same reason companies have departments, not one inbox.
Independent tasks — write tests, draft docs, audit dependencies — fan out to run simultaneously. Wall-clock time approaches the longest single task instead of the sum of all tasks.
The honest caveat: more agents is not automatically better. Every delegation hop loses information ("telephone game"), costs tokens, and adds failure surface. Research on multi-agent framework failures finds most breakdowns are coordination problems, not model problems. A 16-division agency is a library of available specialists — a good orchestrator activates two or three per task, not all fifty.
| Stage | Agent | Context it carries | Returns to orchestrator |
|---|---|---|---|
| 1. Plan | architect | Ticket + repo map (~12k tokens) | 5-step plan, files to touch (~600 tokens) |
| 2. Build | coder | Plan + only the 4 relevant files (~18k tokens) | Diff + rationale (~1.2k tokens) |
| 3. Verify | reviewer | Diff + tests + style guide (~10k tokens) | Pass/fail + 3 findings (~400 tokens) |
| Orchestrator total | Peak ~20k tokens at any moment | vs. ~45k+ for one monolithic thread | |
"An entire AI agency you drop into Claude Code" sounds like magic, but the deliverable is mundane and genuinely useful: a curated folder of subagent definitions plus an installer that copies them into your tool's agent directory. The value isn't secret intelligence — it's that someone already wrote sixteen divisions' worth of tight role prompts so you don't start from a blank page. Evaluate any roster the way you'd evaluate hires: read the prompt files, delete what you'd never use, and measure whether output quality actually improves on your tasks.
The team diagram counts one boss, one node per division, and a chosen number of specialists under every division. Total nodes equal one plus divisions plus divisions times specialists. Four divisions with three specialists each give seventeen nodes. Eight divisions with the same three specialists give thirty-three, shown at twenty pixels per roster node. Doubling divisions does not double the total because the boss is counted only once. Every division connects to the boss, and every specialist connects to its own division. The displayed depth is always two. This visualization creates graph objects, not running software agents, delegated tasks or measured work output. The context readout assigns eight thousand tokens per agent when isolation is enabled. With isolation disabled it assigns eight plus three point two times roster size, in thousands, capped at one hundred eighty. For seventeen nodes, that gives sixty-two point four thousand, shown against eight thousand on a common scale of eight pixels per thousand tokens. The interface rounds the shared figure to sixty-two k, but the underlying formula is sixty-two point four. This is an illustrative formula, not observed prompts, tokenizer output or a real model's context usage. The diagram does not prove that real agents read only their own briefs. Turning isolation off also shows a red haze, which is cosmetic evidence of the chosen toggle rather than measured information interference. With shared context, thirty nodes receive the medium risk label and thirty-one receive high, because the condition is strictly greater than thirty. Their assigned context values are one hundred four and one hundred seven point two thousand tokens, shown at three pixels per thousand. With isolation enabled the label is low through sixty nodes and medium above sixty; high is never assigned in that branch. These categorical rules are not calibrated failure probabilities. Ring radius grows by point two two per division, and the haze scale grows with the square root of total nodes. Those geometry changes do not measure capability or cost. The saved source initializes Three before registering controls, so an unavailable library blocks offline interaction. Native preservation and actual video playback require separate checks; no backend or real agent execution is performed here.