Scheduling & Orchestration

Wunderland now exposes the AgentOS orchestration authoring APIs through wunderland/workflows, with execution routed through createWunderland().runGraph(...) and streamGraph(...).

The important distinction:

• Authoring is code-first today: workflow(), AgentGraph, and mission()
• Execution is in-process through Wunderland’s runtime bridge
• The wunderland workflows CLI command helps discover definitions and examples, but it is not yet a full local workflow runner

Pick the Right Layer

API	Use it when	Strength
workflow()	You know the exact steps and want a DAG	Deterministic pipelines, judge branches, explicit ordering
AgentGraph	You need loops, routers, or custom control flow	Full graph control
mission()	You know the goal but want the planner to decide the steps	High-level intent-driven orchestration

Current CLI Surface

wunderland workflows list
wunderland workflows examples
wunderland help workflows

The list subcommand scans conventional directories such as:

• ./workflows/
• ./missions/
• ./orchestration/

Programmatic Execution

import { createWunderland } from 'wunderland';
import { workflow } from 'wunderland/workflows';

const app = await createWunderland({
  llm: { providerId: 'openai' },
  tools: 'curated',
});

const compiled = workflow('research-pipeline')
  .input({
    type: 'object',
    required: ['topic'],
    properties: {
      topic: { type: 'string' },
    },
  })
  .returns({
    type: 'object',
    properties: {
      finalSummary: { type: 'string' },
    },
  })
  .step('research', {
    gmi: {
      instructions: 'Research the topic and return JSON like {"scratch":{"research":{...}}}.',
    },
  })
  .then('judge', {
    gmi: {
      instructions: 'Return JSON like {"scratch":{"judge":{"score":8,"verdict":"ship","reasoning":"..."}}}.',
    },
  })
  .compile();

const result = await app.runGraph(compiled, { topic: 'agent orchestration frameworks' });
console.log(result);

LLM-as-Judge Pattern

Use a dedicated judge node instead of mixing evaluation into the writing step.

Recommended pattern:

1. Research node writes structured state to scratch.research
2. Judge node writes structured state to scratch.judge
3. Branch on state.scratch.judge.score or state.scratch.judge.verdict
4. Final writer node writes to artifacts

Example judge output:

{
  "scratch": {
    "judge": {
      "score": 8,
      "verdict": "ship",
      "reasoning": "Coverage is broad and the claims are grounded."
    }
  }
}

Chain-of-Thought Guidance

Do not build your orchestration around raw hidden chain-of-thought dumps.

Prefer:

• concise rationale fields like reasoning, score, verdict, risks
• structured JSON the next node can branch on
• explicit artifacts written by the node that owns final output

Avoid:

• prompts that demand the full private reasoning transcript
• branching on unstructured prose

Scheduling

Scheduling still uses the existing cron/job surface. Use orchestration for the graph itself, and cron for when it should run.

wunderland cron list
wunderland cron add "0 9 * * 1-5" daily-report