Anthropic lets one agent launch up to 1,000 sub-agents: 5 repos so the swarm doesn't burn you

Anthropic's managed-agents docs describe a lead agent that can launch up to 1,000 sub-agents over a run's lifetime, with up to 64 concurrent threads, not guaranteed (multi-agent orchestration). We wanted to know what stops a swarm that size from spending your budget before you notice, and whether the tooling you would put in front of it still runs in CI with no API keys. So we re-ran our keyless gate for minilab-paperclip-swarm-board on 2026-10-13, and every number below comes from that run.

The catch

The headline also burns cash. Read the two limits carefully. 1,000 is a lifetime cap for one run, and 64 is a ceiling on concurrency that the platform does not promise to hit. A lead agent told to "use as many workers as you need" will find both limits the expensive way. Prompt-level rules ("be frugal", "stop after ten") do not cap spend. The controls that hold sit outside the model: a session budget, platform limits you have hedged in code, a kill switch that is not an LLM call, and a CI path that stays green without anyone's personal key.

What we ran (keyless)

The lab ships scripts/report.ts. It runs three steps in order, records pass/fail and elapsed milliseconds per step, and writes reports/gate.json plus reports/gate.md:

bun install && bun run report
# = bun run test
#   && bash scripts/upstream-vitest.sh   # pinned Paperclip clone, keyless Vitest subset
#   && bash scripts/paperclip-smoke.sh   # npx paperclipai@latest --help / onboard --help

The upstream step clones paperclipai/paperclip at the pinned commit f329b24ee12496e710ad17377ec7716e0eee22ef, runs pnpm install --frozen-lockfile, then check:node-version, test:check-no-git-push and test:runner-acceptance. It does not run the full monorepo suite (UPSTREAM_FULL_TEST=1 is a deliberate opt-in) and it is not Playwright e2e. The smoke step only prints the CLI help; it never runs onboard --yes.

From reports/gate.json (run on October 13, 2026, Bun 1.4.2, Node 26.3.0):

Two rules keep the report honest. A clone or pnpm install that fails for lack of network is recorded as a failed step with the reason (ok: false, report still written), and test:check-no-git-push prints no TAP summary, so it is recorded by exit code only.

The tools (#5 → #1)

Stars, language and license below were fetched with gh api on October 13, 2026.

#5 · Answer me with HTML

QingYunA/answer-me-with-html · 3,274 ★ · JavaScript · MIT. It returns the answer as a page. Dozens of worker reports are easier to review as one HTML page than as a scroll of markdown.

#4 · Marketing Skills

coreyhaines31/marketingskills · 54,501 ★ · JavaScript · MIT. About fifty installable specialist skills. Give a swarm role a real playbook instead of "act like a marketer".

#3 · notebooklm-py

teng-lin/notebooklm-py · 19,713 ★ · Python · MIT. Cited memory on top of unofficial Google APIs, so prototypes only. Keep its e2e tests out of CI.

#2 · Ponytail

DietrichGebert/ponytail · 162,420 ★ · JavaScript · MIT. Ponytail makes the agent behave like a lazy senior and stop overbuilding, which is cheap with one agent and ruinous with 64.

#1 · Paperclip

paperclipai/paperclip · 101,362 ★ · TypeScript · MIT. An org chart for agents, with budgets and hard stops. The brake sits beside the workers, outside the prompt. Upstream quickstart is npx paperclipai@latest onboard --yes; our smoke only reads --help.

Minilab (open source)

The lab proves one narrow thing. A pinned Paperclip checkout passes its keyless test subset and its CLI answers --help, in under three minutes, with no secrets. It does not prove the full suite is green, that Paperclip caps a real Anthropic run, or the 1,000 / 64 numbers for your account. Re-run bun run report before you quote any of this.

bun install && bun run report