Band of Agents · Track 3 — Regulated & High-Stakes Workflows

WarRoom

Four AI agents — on different frameworks, across two company accounts — run a security incident together through one Band room. A human decides the irreversible calls. The transcript is the audit trail.

View on GitHub → How it works ↓
WarRoom · INC-C · live decision timeline
Triage — BRIEF: critical ransomware on the primary customer DB; recruiting Intel + Compliance.
Threat Intel — FINDING: BlackHaze, lateral movement to domain controllers → isolate + wipe.
Compliance — VETO: host under legal hold; GDPR Art. 33 · ⏱ T-minus 71h 06m. Wipe needs human authorization.
Commander — ESCALATION → @CISO: eradicate vs. preserve. Decision required.
👤 CISO — RULING: preserve evidence, then authorize the wipe. Release the hold.
Commander — RESOLUTION: isolated · imaged · notified · wiped (authorized). Clock running.

A team of agents

Different frameworks. Two organizations. Genuinely different systems negotiating in shared context — a real test of Band as an interoperability layer.

LangGraph

Triage

Classifies the alert and recruits the right specialists into the room.

LangGraph

Threat Intel

Malware attribution and lateral-movement spread assessment.

Pydantic AI · 2nd account

Compliance

External counsel. Owns the live regulatory clock and holds veto power.

Anthropic

Incident Commander

Drives the response; executes actions only after explicit sign-offs.

The scenario: a true dilemma

INC-C — ransomware on the primary customer database (PII + financial data), spreading toward the domain controllers.

Threat Intel says

Isolate and wipe the host now — eradication is the only way to stop the spread before it reaches the domain controllers.

Compliance says

That host is forensic evidence under a legal hold — destroying it is not permitted. A GDPR Art. 33 72-hour clock is now running.

Neither agent is wrong → a genuine deadlock. So the Commander does the only correct thing: it escalates to a human CISO, who rules in a single message. Evidence is preserved first, the wipe is then authorized — and every destructive action stays gated behind a sign-off or a human ruling.

How it works

  1. Alert fires. Triage classifies it and recruits the specialists it needs into one Band room.
  2. Specialists investigate. Threat Intel and Compliance post findings via @mention; asymmetric knowledge forces them to talk.
  3. Deadlock → human. When findings conflict on an irreversible action, the Commander escalates to the human CISO.
  4. Gated execution. Only after the ruling does the Commander act — isolate, preserve evidence, notify, then the authorized wipe.
  5. Audit trail. One command turns the room into a structured incident report — timeline, ruling, regulatory clock, every action.

Why it matters

🔌

Multi-framework interop

LangGraph, Pydantic AI, and Anthropic agents coordinating through one Band room.

🏢

Cross-account

Compliance runs on a separate company's Band account as external counsel.

🧑‍⚖️

Human in the loop

Autonomy where it's safe; human authority for irreversible, regulated calls.

⏱️

Live regulatory clock

GDPR Art. 33 starts a 72h T-minus countdown that rides every message.

🧾

Audit trail by default

The transcript isn't written after the incident — it is the incident.

🛡️

Sign-off protocol

No destructive action runs without a sign-off or an explicit human ruling.