Skip to content

README.md

Latest commit

 

History

History
96 lines (78 loc) · 5.03 KB

README.md

File metadata and controls

96 lines (78 loc) · 5.03 KB
Raw

Local course

Data + Agent Hackathon: hello world, on your laptop · about 45 minutes, plus image downloads

You build an agent whose context is a live Kafka stream, kept fresh by streaming SQL, and that asks a human before it acts. Every part runs on your machine. No account, no team card.

The story

Aegis Financial, a fictional bank, streams every login attempt into Kafka. Somewhere in that stream, an attacker is guessing passwords. Your agent spots them from live data, and flags the account once you say so.

flowchart LR
    K["Ursa for Kafka<br/>diskless topic<br/>security.login_events"] --> S["RisingWave<br/>materialized view<br/>login_failures"]
    J["inject<br/>(you, in Lab 3)"] -- "new login burst" --> K
    S -- "RisingWave MCP<br/>run_select_query" --> G["AI Gateway"]
    G --> A["Orca agent<br/>hello-agent-&lt;you&gt;"]
    A -- "insert_multiple_rows<br/>(only if you approve)" --> F["table<br/>flagged_accounts"]
Loading

What runs on your laptop

Part What it is Started by
Ursa for Kafka Kafka, with a diskless login topic: its records live in an object store, not on the broker docker compose -f local/compose.yaml up
Oxia, object store, schema registry What a diskless topic and Avro need the same command
RisingWave Streaming SQL: the materialized view the same command
RisingWave MCP server The agent's SQL tools the same command
Orca Agent Engine Runs your agent: registry, harness, and AI Gateway local/engine.sh, which runs ork local

The labs

Lab Time Where You The idea
0. Set up 15 min terminal Start both stacks, load the login stream, run the doctor Know that every part answers before you build on it
1. Hello, agent 5 min CLI / Python / TS Create an agent and chat Agent, environment, session, events
2. Hello, streaming SQL 8 min psql Connect RisingWave to the topic and build a materialized view Context that keeps itself fresh
3. Agent + live context 9 min CLI / Python / TS Give the agent a SQL tool, inject new data The answer changes with the data
4. Agent acts, human approves 5 min CLI / Python / TS Let the agent write, with your OK Governed actions

The times are for the steps. Each lab also has a short quiz and a task to try on your own.

What you need

  • Docker with Compose v2 (Docker Desktop, or Docker Engine on Linux). The images are a 5 GB download and take about 20 GB of disk once unpacked; more than half of that is RisingWave. Running, the two stacks use about 2.5 GB of memory.
  • An Anthropic API key, from the Anthropic Console, on an account with credit. The agent answers six questions in the whole course, plus the ones you ask on your own; see "What was run" below for what that cost.
  • ork v0.6.0 or newer, and jq.
  • One path: Python 3.11 or newer, or Node.js 20 or newer. The CLI path also uses one of them for the doctor, the seeder, and the injector.
  • macOS or Linux. On Windows, use WSL 2.

Start with Lab 0: Set up. If something goes wrong, see Troubleshooting. How labs and checks work is in The labs, and a coding agent can tutor you through the course.

Have a team card from the hackathon? Take the Cloud course instead: the same labs, on StreamNative Cloud.

What was run

This course was run on 2 October 2026 on macOS 26 (Apple silicon) with Docker 29.2 and Compose 5.1, from a fresh local/down.sh --reset on each path:

Part Version
ork 0.6.0 (Agent Engine 0.5.1, AI Gateway 0.4.3)
Ursa for Kafka lakestream/kafka:4.3.1.3, with Oxia 0.16.7 and RustFS 1.0.0
RisingWave v3.1.0, with risingwave-mcp-server 0.1.0
Schema registry Karapace 6.2.2
SDKs runorca 0.3.0 on Python 3.13; @runorca/orca-sdk 0.2.3 on Node.js 20
  • Labs 0 and 2: every command and every check, as written, with each check run before its step and after it. Lab 0 on all three paths, each from a fresh local/down.sh --reset.
  • Labs 1, 3 and 4, with the model answering (claude-sonnet-4-6): every step and every check on all three paths. The agent replied, queried the view, saw the injected attack on the second question, proposed an insert, wrote the row when allowed, and stopped when denied. The sample output in Labs 1, 3 and 4 is from those runs, shortened.
  • "Try it yourself": all five tasks, on the Python path (Lab 1 also on the CLI path).
  • What the model calls cost: Labs 1, 3 and 4 taken once, on the CLI path, came to about 310,000 input tokens (three quarters of them read from the prompt cache) and 2,500 output tokens: about $0.35 at Sonnet's list prices. Each turn sends the model about 30,000 tokens of context.