Band of Agents Hackathon · Track 3 — regulated & high-stakes

Insurance claims, adjudicated by a band of agents.

Four specialist AI agents — each on a different framework and model vendor — collaborate inside one Band room to triage an auto-insurance claim and decide APPROVE, DENY, or ESCALATE to a human.

IntakeLangGraphCoverageGemini SDKFraudLangGraphAdjudicatorCrewAIhuman reviewer
4
remote agents
3
frameworks
2
model vendors
1
Band room
Watch a claim get adjudicatedAdjudicate your own claim

Live relay · real captured runs

Watch a claim move through the band

Pick a claim. Each agent reads the shared record from the Band room, runs its logic, and hands off to the next. Where rules fall short, its model reasons over the claim — watch the Fraud agent read the incident storyand add risk the rules can't see. Captured from a live run.

CLM-CLEAN→ APPROVEA valid, low-risk collision claim.
claimant
Jane Doe
vehicle
2021 Honda Civic
incident
collision · 2026-06-10
claimed
$4,200
policy
active
  1. 01

    Intake

    LangGraph · Groq

    • valid
    • completeness 100%
    • 0 missing · 0 conflicts
  2. 02

    Coverage

    Gemini SDK · Gemini

    • policy active
    • peril covered
    • covered $3,700

    • Reasoning

      • 1. The policy is active for the incident date (2026-06-10 falls between 2026-01-01 and 2026-12-31).
      • 2. The incident type is collision, which is covered by the policy.
      • 3. The coverage amount is calculated using the formula: min(estimate, limit) - deductible.
      • 4. min($4,200 (estimate), $50,000 (limit)) = $4,200.
      • 5. Subtracting the deductible: $4,200 - $500 = $3,700.
  3. 03

    Fraud

    LangGraph · Groq

    0/100 risk

    0 rules

    “Consistent rear-end collision damage and claim amount.”

    Analysis

    • 1. Police report was filed.
    • 2. Claimant has no prior claims in the last 12 months.
    • 3. Sufficient photos provided (6).
    • 4. Claimant is the policy holder.
    • 5. Claim amount ($4,200.00) is below the high-risk threshold.
    • 6. Incident occurred 160 day(s) after policy effective date.
    • 7. Calculated risk score: 0 based on 0 triggered red flag(s).
    • 8. LLM narrative judgment (+0): Consistent rear-end collision damage and claim amount.
  4. 04

    Adjudicator

    CrewAI · Groq

    verdict →

    APPROVE

    The policy is active on the incident date, the fraud risk score is low (0) and the covered amount ($3,700) is well below the escalation threshold. All hard rules point to approval, so the claim is approved for the covered amount.

Final decision · CLM-CLEAN

APPROVE

Payable

$3,700

The policy is active on the incident date, the fraud risk score is low (0) and the covered amount ($3,700) is well below the escalation threshold. All hard rules point to approval, so the claim is approved for the covered amount.

The Band room trail

The actual messages from the live room — the full audit trail. The verdict is broadcast to every agent and the human reviewer.

  1. 13:56:08Coverage@intake New claim submitted for review: Please validate.
  2. 13:56:36IntakeIntake Agent — Claim intake is valid with no missing fields or inconsistencies. Handing off to @coverage for the next step.
  3. 13:56:49CoverageCoverage Agent — Policy active, peril covered; payout is $3,700 after $500 deductible. Handing off to @fraud for the next step.
  4. 13:56:54FraudFraud Agent — Claim appears legitimate with no red flags and a low risk score of 0. Handing off to @adjudicator for the next step.
  5. 13:56:59Adjudicator@fraud @coverage @intake @nivishnick2k (human) Adjudicator Agent — Claim approved for $3,700, policy is active and fraud risk is low. Final decision: APPROVE.

Deterministic guardrails · runs in your browser

Adjudicate your own claim

Edit the claim and run the deterministic policy guardrailsthe agents must obey — live in your browser. These set the hard floor a regulated workflow can't cross. On top of them, the agents' models add narrative judgmentthe rules can't make — e.g. the Fraud agent reading the incident story (see the replay above).

Start from
Policy
Incident & claimant

Why this wins: real cross-framework collaboration

Most agents work alone. ClaimBand puts three frameworks and two vendors in one room and makes them finish a regulated workflow together.

Band is the collaboration layer

Agents exchange the full claim record as shared room context, route work with @mentions, and broadcast the verdict to the whole band — not a thin wrapper around a final notification.

Rules as guardrails, models for judgment

Data flows deterministically through Band's shared context (reliable — no JSON-shuttle failures). On top, each agent's model makes calls the rules can't: the Fraud agent reads the incident narrative and adds risk a field-check would miss. The deterministic guardrails set the hard floor; the model decides within them.

Resilient by design

When Gemini's free tier hits its 20/day cap and returns 429, the Coverage agent falls back to a templated note and the relay still completes end-to-end. No data loss, no crash.

Acceptance criteria — honestly

Five of six pass against the live runs. The sixth is wired in code and waiting on a fresh live room capture; a credible system names its edges.

  • CLM-CLEAN → APPROVE, correct amount$3,700 payable
  • CLM-DENY → DENY with reasonpolicy expired
  • CLM-FRAUD → ESCALATE + human @mentionrisk 60, human-in-loop
  • Different frameworks AND vendors at startupshown in startup banner
  • No agent crash, WebSocket reconnect handled3/3 runs clean
  • Band peer-discovery on ambiguous-risk pathwired in relay, live capture pending