agent-echo

The smallest LLM-backed experience: one agent node reads a name from args and writes a one-sentence greeting.

What it demonstrates

• A single agent node with a prompt file
• Prompt interpolation: resolved from node args
• ANTHROPIC_API_KEY as the minimal LLM requirement
• Structured output via the agent's writes contract

The graph

graph:
  nodes:
    - id: greet
      kind: agent
      prompt: ./prompts/echo.md
      args:
        name: World
      writes: [greeting]
  edges: []

State schema

Field	Type	Direction	Description
name	string	in (via node args)	Who to greet
greeting	string	out	Model's one-sentence reply

How it runs

Requires ANTHROPIC_API_KEY:

export ANTHROPIC_API_KEY=sk-...
oe run examples/agent-echo

To greet a specific name:

oe run examples/agent-echo --args '{"name":"Alice"}'

LLM required

This is the first example that makes a real API call. Make sure your ANTHROPIC_API_KEY is set before running.

What happens

1. The scheduler dispatches greet to AgentDispatcher.
2. The dispatcher loads prompts/echo.md, interpolates → "World" (or whatever was passed).
3. The model receives: "Say a friendly hello to World. Keep it short — one sentence."
4. The model's text response is written to the greeting field in state.

$ oe state greeting
Hello, World! Wishing you a wonderful day ahead.

The prompt

Say a friendly hello to {{name}}. Keep it short — one sentence.

Prompt files live in prompts/ and use interpolation. This one is the simplest possible: a single sentence instruction with one variable.

Try it: variations

1. Change the prompt. Edit prompts/echo.md to ask for a different style — e.g., "Greet in the style of a pirate." No schema changes needed.

2. Add a second agent. Add a farewell agent node that reads greeting and writes goodbye. The node can reference in its prompt. Add edge { from: greet, to: farewell }.

3. Use a different model. By default the AgentDispatcher uses whatever model your client is configured for. Pass --model claude-3-5-haiku-20241022 (or the equivalent CLI flag) to use a faster/cheaper model for this toy example.

Compare with hello-tool

hello-tool does the same thing without a model call. Compare the two to understand exactly what the agent node kind adds.

Source

examples/agent-echo/