Getting Started - Plan-Execute From YAML¶
The plan_execute_from_yaml.py workflow is useful for longer tasks that benefit from explicit planning, execution, checkpointing, and restart prompts.
Current status
This workflow currently lives in the examples directory. It is useful today for long, complex jobs, and future URSA versions are expected to integrate this workflow more directly into the main URSA interface.
Prerequisites¶
- URSA is installed.
- You have access to an LLM endpoint.
- You are running from a clone of the URSA repository or otherwise have the
examples/two_agent_examples/plan_execute/files available. - You have a dedicated workspace for generated files.
1. Choose an example YAML file¶
The repository includes example task files under:
examples/two_agent_examples/plan_execute/
A good first run is:
examples/two_agent_examples/plan_execute/example_from_yaml.yaml
Other examples include pi_multiple_ways.yaml, openchami_boot_docs_example.yaml, and larger scientific demos.
2. Run the workflow¶
python examples/two_agent_examples/plan_execute/plan_execute_from_yaml.py \
--config examples/two_agent_examples/plan_execute/example_from_yaml.yaml \
--workspace ./plan-execute-workspace
The runner may ask you to:
- select or confirm an LLM model,
- choose single or hierarchical planning,
- resume from a checkpoint or start fresh.
For a first run, choose the simplest/single-planning option and start fresh.
3. Review the generated plan¶
The planning stage decomposes the YAML task into steps. Read the plan before allowing a long workflow to proceed, especially if the execution step may write files, install packages, or run commands.
4. Review outputs in the workspace¶
The workspace contains generated code, outputs, checkpoints, and artifacts created during the run. Treat it as the durable record of the job.
5. Resume or inspect a run¶
For detailed checkpointing, resume behavior, and troubleshooting, see the Plan-Execute checkpointing reference.