Skip to content
Azhi FlowAzhi FlowAzhi Flow

Write a workflow

A workflow is a package: a folder with workflow.yaml and the files it references (profiles, schemas, templates, scripts, knowledge). You can write it by hand, start from a template, or build it with chat.

  1. Start from a template.

    Terminal window
    npx azhi init my-report                # the weekly quality report (default)
    npx azhi init my-digest -t ci-digest
    npx azhi init my-delivery -t sdlc      # feature delivery with design and release gates

  2. Validate locally. Compiles the package and prints diagnostics: graph, ref types against output schemas, tool arguments and projections, CEL, taint.

    Terminal window
    npx azhi validate my-report
    npx azhi validate my-report -c azhi.config.yaml    # also checks the tools
    npx azhi schema                                    # print the workflow JSON Schema (2.x)

  3. Register tools and read the plan.

    Terminal window
    npx azhi apply my-report/azhi.config.yaml
    npx azhi plan my-report

  4. Test one node on a fixture, with writes mocked.

    Terminal window
    npx azhi test-node metrics my-report -f fixtures/metrics.json

  5. Run it.

    Terminal window
    npx azhi run my-report -i team=payments --wait
    npx azhi inspect <run-id>

  6. Publish and schedule. A published version’s trigger.schedule becomes the workflow’s schedule, in an IANA timezone.

    Terminal window
    npx azhi publish my-report
    npx azhi schedule my-report --cron "0 8 * * 1" --timezone Asia/Kolkata

Anatomy

Section titled “Anatomy”
schema_version: "2.0"
id: quality-report
name: Weekly quality report
trigger:
  schedule: {cron: "0 8 * * 1", timezone: "Asia/Kolkata"}
  manual: true
inputs:
  type: object
  properties: {team: {type: string}}
  required: [team]
config:
  channel: C-QUALITY
nodes:
  # ... tool → script → agent → report → notify
  - id: post
    type: notify
    channel: slack
    destination: {ref: config.channel}
    message: {ref: nodes.report.output.markdown}
    guard: "args.channel == config.channel"   # taint gate: analyse read untrusted data

Habits that pay off

Section titled “Habits that pay off”
HabitWhy
Put numbers in a script and let the agent explain themAuthoritative figures, fewer tokens
Use project: on tool nodesThe agent sees the fields it needs, not the whole payload
Give every agent an output_schema and a budgetValidated output (with two repairs) and a spend cap
Gate every write after an agent with a guard or an approvalOtherwise it will not compile
Run azhi plan in CICatches an unsatisfiable workflow before a scheduled run does
Pin datasets by tag (name@approved)Reproducible knowledge