Skip to main content

​
1. Pick a runtime

Browse Runtimes and choose by capability. For a first workflow, anthropic-agent-sdk (agentic coding) or gemini-genai-sdk (research/chat) are good defaults — both with method: sdk.

​
2. Write a minimal workflow

Create a workflow.md with frontmatter naming your runtime, a budget, tools, and a sandbox. Keep it declarative — policy, not code: 
---
id: my-first-workflow
version: 1.0.0
agent:
  runtime_key: anthropic-agent-sdk
  provider: anthropic
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
  network:
    - api.anthropic.com
tools:
  required:
    - fs.read
    - fs.write
    - git
---

# My First Workflow

Plan, implement, verify, and record evidence before finishing.

​
3. Validate

Validate the workflow before dispatch — Cadence rejects invalid policy before anything runs (and your editor agent can do this for you via the think_workflow_md_validate tool).

​
4. Dispatch through Cadence

Dispatch the workflow from the deeda Think app or via the harness tools. Never bypass Cadence with a raw provider call for production work: budgets, sandboxing, retries, review, and evidence all live on the Cadence path.

​
Next steps

Workflow Markdown

The full schema surface: selection, fallbacks, review consensus, state machines.

Capability Catalog

Everything each provider can do through the harness today.