Band of Agents / ClaimBand
A four-agent insurance-claim prototype that carries validated Pydantic state across three frameworks and preserves an inspectable approve, deny or escalate trail.
Problem
A multi-agent decision is difficult to trust when hand-offs and model reasoning disappear between steps. The prototype needed a shared context and an inspectable decision trail.
How I solved it
- 01
Carries one validated Pydantic claim record through a four-agent relay instead of reconstituting unvalidated text at each hand-off.
- 02
Combines LangGraph, the Gemini SDK and CrewAI around shared Band-room context so coverage, fraud and escalation decisions can be followed across framework boundaries.
- 03
Keeps hard coverage, fraud and escalation rules deterministic and captures decision trails, so a provider failure or model output cannot silently override a guardrail.
My role / what I owned
Hackathon builder
- Built the four-agent relay across LangGraph, Gemini SDK and CrewAI.
- Kept one validated Pydantic claim record through shared Band-room hand-offs.
- Added deterministic coverage, fraud and escalation guardrails plus captured evidence trails.
How it was built
- Team
- A Band of Agents Hackathon entry; the public project does not document a human collaborator split.
- Tools
- LangGraph
- Gemini SDK
- CrewAI
- Groq
- Pydantic
- Band shared context
- Use of AI
- Models provide specialist judgement and narration, while deterministic rules preserve coverage, fraud and escalation boundaries when a provider fails or violates a hard rule.
Source and current state
- 14 public commits at the 9 August 2026 review.
- Seven direct test modules plus fixtures, captured decision trails and submission artefacts.
- Status
- Seven-day hackathon prototype with public source and artefacts.
- What it shows
- The prototype demonstrates traceable orchestration across four agents and three frameworks.