Summary
Allow agents to reference external workflow files as sub-workflows via a new type: workflow agent type, enabling composable and reusable workflow building blocks.
Motivation
The industry is moving toward composable agent architectures — Stripe's "blueprints" (deterministic routines composed with LLM loops), Google ADK's progressive disclosure (agents loading skills on demand), and Claude Managed Agents with multi-agent coordination all point to the need for workflow composition.
Today conductor workflows are flat single-YAML files. Complex workflows become unwieldy and can't share common patterns across projects.
Proposed Design
agents:
- name: deep_research
type: workflow # new agent type
workflow: ./research-pipeline.yaml # reference to external workflow
input:
topic: "{{ workflow.input.topic }}"
output:
findings:
type: string
routes:
- to: synthesizer
Key Semantics
- Sub-workflow runs as a black box — internal agents aren't visible to the parent
- Input/output contract enforced at the boundary
- Sub-workflow inherits parent's provider config unless it specifies its own
- Recursive composition supported (with depth limit for safety)
- Sub-workflow events bubble up to parent's event system (prefixed with sub-workflow name)
Validation
conductor validate follows sub-workflow references and validates the full tree
- Circular references detected and rejected
- Input/output type compatibility checked across boundaries
Why It Fits Conductor
- YAML-native, version-controllable, no new runtime concepts
- Clean separation of concerns enables a library of reusable workflow "skills"
- The sub-workflow is just another YAML file — same tooling, same validation, same dashboard
Effort Estimate
Medium — requires new agent type in executor, schema changes, and cross-file validation.
Summary
Allow agents to reference external workflow files as sub-workflows via a new
type: workflowagent type, enabling composable and reusable workflow building blocks.Motivation
The industry is moving toward composable agent architectures — Stripe's "blueprints" (deterministic routines composed with LLM loops), Google ADK's progressive disclosure (agents loading skills on demand), and Claude Managed Agents with multi-agent coordination all point to the need for workflow composition.
Today conductor workflows are flat single-YAML files. Complex workflows become unwieldy and can't share common patterns across projects.
Proposed Design
Key Semantics
Validation
conductor validatefollows sub-workflow references and validates the full treeWhy It Fits Conductor
Effort Estimate
Medium — requires new agent type in executor, schema changes, and cross-file validation.