RunSybil's November post on multi-agent orchestration put a number on what this work actually takes: ~2,700 coordinated agents per engagement, sharing real-time context like an elite human team. The point isn't the headcount — it's the correlation. Most critical findings aren't bugs; they're relationships between bugs.
Their example is instructive: a low-priority file-read vulnerability, alone, is a shrug. Correlated with a session-signing mechanism discovered elsewhere in the app, it yields the signing key — and a forged administrative session token. Two findings, one takeover. Neither agent saw it alone.
Why chains need a shared brain
Independent agents produce independent findings. A swarm that shares context produces something else: a lattice where "host A leaks X" meets "service B trusts X" and a chain snaps into existence. Our orchestration works the same way — findings land in a shared context bus where chain agents hunt for combinations:
# chain hypothesis — swarm internal finding[041]: review-app-2214 file-read /app/.env severity:low finding[118]: api.internal jwt-rs256 kid accepts path severity:info → chain.07 proposes: read signing key via 041 → forge admin jwt via 118 → executed, verified: full admin session. severity:critical
The operator's job in the swarm era
Humans aren't replaced by orchestration — they're promoted. The swarm does breadth at machine speed; the operator reviews proposed chains, kills dead ends, and pushes into the places autonomy still can't reach: business logic, chained context, the creative cruelty of a real intruder. Every report that ships carries a human signature, because someone has to stand behind "proven."
Single findings are inventory. Chains are intrusions. If your pentest report is a list of severities and not a list of demonstrated paths, you bought the first kind.