Steward AI

AI Hackathon 2026/Architect & builder
Steward AI demo

AI Hackathon 2026 project for building agent-native systems where agents are first-class citizens. A supervisor routes questions through a Planner / Executor / Reviewer loop over SME sub-agents on A2A, with streaming text and generative UI in an embeddable widget.

4
agents: supervisor + 3 SMEs
10
generative UI components
P/E/R
planner / executor / reviewer
A2A
SME discovery over HTTP
Role
Architect & builder
Timeline
AI Hackathon 2026
Stack
React 19, TypeScript, Vercel AI SDK, Hono, LangGraph, Strands A2A, AWS Bedrock AgentCore, Terraform, Aurora PostgreSQL, CloudFront
Tags
AI Agents, Multi-Agent, Generative UI, AI Hackathon 2026

§Context

This was an AI Hackathon 2026 full of learnings for agent-native systems: agents as first-class citizens, not a thin chat wrapper on one model. The goal: one embeddable conversational workspace where property-ops questions cross leasing, communications, and make-ready without the user jumping tools.

§What we built

  • Embeddable React widget (Shadow DOM) with split chat + workspace panel, streamed through the Vercel AI SDK.
  • Supervisor gateway that either replies directly or hands off to an orchestrator with a Planner / Executor / Reviewer loop (max three iterations) over multiple SME sub-agents.
  • Three domain SMEs (Lease Management, Communications, Property Operations) called over the A2A protocol; new agents discover via agent cards, not a hardcoded registry.
  • Generative UI pipeline: agents emit UI directives; the proxy translates AgentCore SSE into AI SDK parts; the frontend lazy-loads tables, charts, kanban, bento, job cards, and more.
  • Async job demo path with portal callback, live JobStatusCard polling, and notification bell.
  • Full AWS shape via Terraform: CloudFront dual-origin, Lambda proxy, Bedrock AgentCore runtimes, Aurora Serverless v2.

§Design choices

  • Thin proxy, fat agents: Hono only bridges AI SDK ↔ AgentCore SSE so the UI never holds secrets or provider logic.
  • Three-layer hierarchy: supervisor for routing and memory; orchestrator for plan-execute-review; SMEs for domain tools only.
  • Schema parity end-to-end: Pydantic in agents, shared TypeScript types in the proxy, Zod in the UI, so UI directives stay honest.
  • Independent executor steps run in parallel; title generation and UI rendering run concurrently so text reaches the user first.
  • Infra under SCP constraints: ALB (900s idle) instead of Function URLs, RDS Data API instead of NAT, default VPC.

§Outcome

A working AI Hackathon 2026 demo: ask about move-ins, watch supervisor to planner to executor to reviewer phases, get text plus interactive UI in the workspace, fire an async job, and see the portal callback land on a live job card. The durable takeaway is the pattern set: A2A discovery, stream translation, generative UI directives, and production-shaped IaC from day one.