go-further.md

Go further

You finished the hello world: live data flows from Kafka through streaming SQL into an agent that can act, with a human in the loop. Here is where to take it during the build session. Items marked ask onsite depend on features whose availability in your environment an onsite StreamNative engineer can confirm.

This page is written for the Cloud course. On the Local course the Agent Engine features below work the same way; the preloaded topics and the StreamNative MCP tools are specific to StreamNative Cloud.

Build ideas

• Richer context. Join the other preloaded topics (identity_changes, threat_intel, device_signals, payment_events) into one materialized view that scores each account. The agent's context stays fresh with no extra code.
• Your own data. Generate a live stream for your idea with ShadowTraffic (covered after the tutorial), register an Avro schema for your topic, and point the same pattern at it.
• Decisions as events. Instead of a SQL table, let the agent write its decisions to a Kafka topic that other systems consume. The StreamNative MCP server offers kafka_client_produce on a cluster route. Ask onsite.
• Event-driven agents. Start a session automatically for every alert instead of when you type: Agent Engine triggers can open one session per Kafka event. Ask onsite.

Agent Engine features to reach for

Custom tools: your code as a tool

Custom tools run in your process, so they can call any API you have credentials for (post to Slack, open a ticket, call your service). Declare one in the agent's tools:

{
  "type": "custom",
  "name": "notify_oncall",
  "description": "Page the on-call analyst about an account. Returns a ticket id.",
  "input_schema": {
    "type": "object",
    "properties": { "account_id": { "type": "string" }, "summary": { "type": "string" } },
    "required": ["account_id", "summary"]
  }
}

When the agent calls it, the session emits agent.custom_tool_use and goes idle with stop_reason.type = "requires_action". Run your code, then answer with:

{
  "type": "user.custom_tool_result",
  "custom_tool_use_id": "<the agent.custom_tool_use event id>",
  "content": [{ "type": "text", "text": "ticket OPS-123 opened" }]
}

The turn loops in this repo (run_turn / runTurn) answer approvals for MCP tools; extend them with a branch for agent.custom_tool_use that calls your function and sends this event. Answer every blocked id, or the session waits forever.

Permission policies

always_allow runs a tool immediately; always_ask pauses the session for a human, as in Lab 4. Keep write tools on always_ask until you trust them, and keep tools you don't need disabled (see default_config.enabled: false in agent/cloud/l4-act.json).

Skills and guardrails

Skills package instructions and reference files (a SKILL.md bundle) the agent reads on demand, for example an investigation playbook. Guardrails cap tool calls or cost per session. See the Orca documentation for both.

Other clients

Orca implements the Managed Agents API, so clients built for that API can talk to your Agent Engine by changing their base URL. Ask onsite before relying on it: authentication details differ between clients.

What makes a strong demo

• It runs live, on the real stack, in front of the judges.
• It needs both halves: real-time data and an agent. If a nightly batch job would do, the streaming half isn't pulling its weight.
• It teaches something about the problem, the data, or the tools.
• Someone else could run it from your README.
• It tells a story in five minutes: the problem, the moment it's solved, the payoff.