Skip to content

demo-runbook.md

Latest commit

 

History

History
95 lines (66 loc) · 3.59 KB

demo-runbook.md

File metadata and controls

95 lines (66 loc) · 3.59 KB

Demo Runbook

Operator's guide for running the 3-minute hackathon demo. Use when rehearsing or on demo day.

Pre-Demo Checklist

  • Program deployed on devnet — solana program show <PROGRAM_ID> --url devnet
  • Server running — curl https://<server-url>/webhook (should return 405)
  • Dashboard live on Vercel — open in browser
  • Neon database has tables — cd server && npx prisma studio
  • Helius webhook configured and pointing to server URL
  • Demo keypairs funded — solana balance <KEYPAIR> --url devnet (need ~5 SOL each)
  • Demo policy created — cd dashboard && npm run demo:setup
  • Backup demo recording uploaded (in case of live failure)

Running the Demo

Terminal 1: Start the simulation

cd dashboard
npm run demo:simulate

This runs:

  • T+0s: Starts trader-agent (Jupiter swaps, ~30 txns/hour) + staker-agent (Marinade)
  • T+60s: Starts attacker agent (burst swaps, unfamiliar programs, drain attempt)

On the Dashboard

  1. Open /agents — show three agents running, all green
  2. Open one agent detail — show spend gauge ticking up, activity feed streaming
  3. Wait for attacker to trigger (~T+65s):
    • First suspicious txn appears yellow (FLAG) in activity feed
    • Claude's reasoning visible inline: "new program + 3x median amount + 4 txns in 8s"
    • Second txn triggers red (PAUSE) — agent stops instantly
  4. Click into /incidents — show the paused incident
  5. Open incident detail — show Opus-generated postmortem with timeline

Expected Timing (ideal conditions — Claude API <2s latency)

Time Event
T+0s Trader + staker start, normal activity streams in
T+60s Attacker starts burst-swapping
T+62s First FLAG verdict from Claude
T+63s PAUSE verdict — on-chain pause executed
T+65s Incident appears in dashboard (via SSE)
T+70s Opus incident report generates (via SSE)

Note: If Claude API latency spikes >3s, the pause may take longer. Do dry runs to calibrate timing before demo day.

Demo Script (What to Say)

0:00–0:20 — Hook: "Step Finance lost $40M because a treasury wallet with full permissions was compromised. Today, 250,000 AI agents operate on Solana with the same risk."

0:20–0:50 — What we built: "Agent Guardrails Protocol — an on-chain policy layer. Three enforcement layers: allow-listing, budget, kill switch." Show architecture diagram.

0:50–1:30 — Live demo: Dashboard on screen. Three agents, activity feed, spend gauges. Everything green.

1:30–2:15 — Attack: Attacker starts. First txn flagged yellow. Claude reasoning inline. Second txn: PAUSE. Agent stops. On-chain event in explorer.

2:15–2:40 — Incident view: Opus postmortem. Timeline. Judge reasoning chain.

2:40–3:00 — Close: "Built on Swig session keys, Squads multisig, Helius streams. Open source, devnet-live today."

Fallback Procedures

Devnet is slow/congested

  • Switch to the pre-recorded demo video
  • Or use localnet: run a local validator and adjust RPC URLs

Server is down

  • Check server logs on your hosting platform
  • Restart or redeploy the server

Claude API latency > 3s

  • Server falls back to rule-based verdict (if implemented)
  • Narrate: "In production, the 3-second timeout triggers a rule-based fallback"

Dashboard won't load

  • Show Prisma Studio (npx prisma studio) as proof of data in the database
  • Show Solana Explorer for on-chain events

Demo agents fail to connect

  • Verify devnet SOL balance: solana airdrop 5 --url devnet
  • Verify program ID matches in demo script env