ADK Dynamic Workflows

Model-generated, validated workflows that can execute against Google ADK 2 or Temporal without evaluating arbitrary generated Python.

This project is an early bootstrap. The implemented vertical slice includes:

• A strict, versioned WorkflowSpec suitable for structured LLM output
• An ADK planner agent backed by LiteLLM structured output
• Agent, sequence, parallel, map, condition, and bounded-repeat steps
• Stable execution IDs for replay and caching
• A backend-neutral asynchronous interpreter
• An ADK 2 adapter built on Context.run_node()
• A Temporal adapter that dispatches agent calls as child workflows

Install

uv sync

Install the Temporal adapter dependencies when needed:

uv sync --extra temporal

The official temporalio[google-adk] extra currently requires google-adk<2. This project therefore uses Temporal's core SDK and leaves the ADK 2 child-workflow implementation as the next integration milestone.

Validate A Workflow

uv run adk-dynamic-workflows validate examples/review.json

Generate A Workflow With LiteLLM

Edit the gitignored .env file with your OpenAI or OpenAI-compatible provider:

DYNAMIC_WORKFLOWS_LLM_MODEL=openai/gpt-4o-mini
DYNAMIC_WORKFLOWS_LLM_API_BASE=https://api.openai.com/v1
DYNAMIC_WORKFLOWS_LLM_API_KEY=replace-with-your-api-key

For a compatible gateway, retain LiteLLM's openai/ model prefix and replace the model name and base URL with values supported by that gateway.

Generate, review, approve, and save a workflow:

uv run adk-dynamic-workflows plan \
  "Review every file provided in the files input" \
  --agent-profile explorer \
  --agent-profile reviewer \
  --output generated-review.json

The planner uses ADK LlmAgent.output_schema to request a WorkflowSpec, then validates the result locally. It retries once with validation feedback when the provider returns a structurally invalid workflow. The generated workflow may only reference profiles passed through --agent-profile.

Execute Locally

Register the agent profiles that a generated spec is allowed to invoke, then build a normal ADK workflow:

from adk_dynamic_workflows.adk_backend import build_adk_workflow
from adk_dynamic_workflows.spec import WorkflowSpec

spec = WorkflowSpec.model_validate_json(open("examples/review.json").read())
root_agent = build_adk_workflow(
    spec,
    {
        "reviewer": reviewer_agent,
    },
)

The allowlisted registry is intentional: generated workflows select an agent profile but cannot create unrestricted agents or grant themselves new tools.

Development

uv run ruff format --check .
uv run ruff check .
uv run pyright
uv run pytest

Run the real LiteLLM integration test after configuring .env:

uv run pytest -m integration tests/integration/test_litellm_planner.py

With placeholder credentials the integration test is skipped. With configured credentials it performs a real provider request and verifies that the returned workflow is valid and uses only the requested agent profile.

Next Milestones

1. Implement an ADK 2-compatible Temporal child workflow for one agent run.
2. Add budget estimates, progress queries, and cancellation.
3. Store large outputs as artifact references instead of workflow-history data.
4. Add semantic validation for branch scope and step-output dominance.