Skip to main content

Workflow Dependencies

Workflow dependencies enable orchestration across multiple workflows, creating multi-stage pipelines.

Dependency Patterns​

PatternBehavior
SingleOutput becomes input
Parallel BranchesOutputs merged into combined records
Fan-inAll available, matched by lineage
AggregationAll outputs merged and grouped by reduce_key

Single Dependency​

- name: validate_data
  dependencies: extract_data

Parallel Branches​

Created using versions - multiple executions of the same action that get merged downstream:

# Step 1: Create parallel branches with versions
- name: research
  dependencies: [analyze_issue]
  intent: "Research the issue from different angles"
  versions:
    range: [1, 3]      # Creates research_1, research_2, research_3
    mode: parallel
  schema: issue_resolution/research_findings
  prompt: $issue_resolution.Research_Issue
  context_scope:
    observe:
      - analyze_issue.*
      - source.*

# Step 2: Consume parallel branches - outputs are MERGED
- name: synthesize
  dependencies: [research]  # Automatically resolves to [research_1, research_2, research_3]
  intent: "Synthesize research findings"
  context_scope:
    observe:
      - research.*  # All branches merged and available

Fan-in (Multiple Different Actions)​

Different actions converging - all available matched by lineage:

# Three different analysis actions run in parallel
- name: analyze_sentiment
  dependencies: [extract_data]
  # ...

- name: analyze_entities
  dependencies: [extract_data]
  # ...

- name: analyze_topics
  dependencies: [extract_data]
  # ...

# Fan-in: all three available, matched by lineage
- name: generate_report
  dependencies: [analyze_sentiment, analyze_entities, analyze_topics]
  context_scope:
    observe:
      - analyze_sentiment.*
      - analyze_entities.*
      - analyze_topics.*

The first dependency determines execution count — order your dependencies list so the action that controls record count appears first.

Aggregation​

Merge ALL outputs from different actions and group by a key. Use reduce_key:

# Three validators each produce validation results
- name: validator_grammar
  dependencies: [generate_content]
  # ...

- name: validator_accuracy
  dependencies: [generate_content]
  # ...

- name: validator_style
  dependencies: [generate_content]
  # ...

# Aggregation: merge ALL outputs, group by content_id
- name: aggregate_validations
  dependencies: [validator_grammar, validator_accuracy, validator_style]
  reduce_key: content_id  # Groups all validations for same content
  context_scope:
    observe:
      - validator_grammar.*
      - validator_accuracy.*
      - validator_style.*

Fan-in vs Aggregation:

  • Fan-in (no reduce_key): First dep drives execution, others matched by lineage
  • Aggregation (reduce_key set): All outputs merged, grouped by key
Auto-Inferred Dependencies

Actions referenced in context_scope but not in dependencies are automatically available via lineage matching.

Cross-Workflow Dependencies​

When a single workflow isn't enough, you can chain workflows together. A downstream workflow declares a dependency on an upstream workflow's output — Agent Actions handles the artifact linking automatically.

How It Works​

When you run with --upstream or --downstream, Agent Actions:

  1. Scans all workflows in agent_workflow/ to build a dependency graph
  2. Resolves execution order using topological sort
  3. Runs workflows in dependency order
  4. Writes .upstream_manifest.json in the downstream workflow's agent_io/ directory, pointing to the upstream workflow's latest output
  5. The downstream workflow reads from the upstream output as its input

Declaring Cross-Workflow Dependencies​

Reference another workflow in your action's dependencies list using the workflow key:

# Workflow B: depends on Workflow A's output
actions:
  - name: enrich_data
    dependencies:
      - workflow: data_preparation        # All terminal outputs from Workflow A
    prompt: |
      Enrich this validated data...

To depend on a specific action's output:

actions:
  - name: enrich_data
    dependencies:
      - workflow: data_preparation
        action: validated_output          # Only this action's output
    prompt: |
      Enrich: {{ validated_output.content }}

Project Layout​

Cross-workflow dependencies require workflows to live under the same agent_workflow/ root:

agent_workflow/
├── data_preparation/          # Workflow A (upstream)
│   ├── agent_config/
│   │   └── data_preparation.yml
│   └── agent_io/
│       ├── staging/           # A's input
│       └── target/            # A's output
│
└── content_generation/        # Workflow B (downstream)
    ├── agent_config/
    │   └── content_generation.yml
    └── agent_io/
        ├── .upstream_manifest.json  # Auto-generated link to A's output
        ├── staging/
        └── target/

The Upstream Manifest​

When Agent Actions links workflows, it writes .upstream_manifest.json into the downstream workflow's agent_io/ directory:

{
  "upstream_workflow": "data_preparation",
  "upstream_path": "/project/agent_workflow/data_preparation/agent_io/target/validated_output",
  "files": ["batch_001.json", "batch_002.json"]
}

The downstream workflow reads this manifest to locate its input data instead of looking in staging/.

CLI Execution​

# Run upstream workflows first, then this workflow
agac run -a content_generation --upstream

# Run this workflow, then all downstream workflows
agac run -a data_preparation --downstream

# Full chain: upstream → this → downstream
agac run -a middle_workflow --upstream --downstream
CommandWhat Executes
agac run -a BB only
agac run -a B --upstreamA → B
agac run -a B --downstreamB → C
agac run -a B --upstream --downstreamA → B → C
Topological Ordering

When multiple workflows have interdependencies, Agent Actions performs a topological sort to determine the correct execution order. Circular dependencies are detected and reported as errors.

Field References​

Reference upstream workflow outputs in prompts just like any other action:

- name: enhance_content
  dependencies:
    - workflow: quiz_generation
      action: format_quiz_text
  prompt: |
    Enhance: {{ format_quiz_text.question }}

See Also​