AI Native Lang

Hybrid deployments: AINL inside Temporal

Temporal gives you durable execution: retries, timeouts, visibility, and long-running workflows across process restarts. AINL gives you a small, deterministic program (graph IR) for policy and adapter logic without encod

Hybrid deployments: AINL inside Temporal

What this is

Temporal gives you durable execution: retries, timeouts, visibility, and long-running workflows across process restarts. AINL gives you a small, deterministic program (graph IR) for policy and adapter logic without encoding that behavior in imperative workflow code.

The --emit temporal option in scripts/validate_ainl.py generates two Python modules:

  1. *_activities.py — defines a Temporal activity that calls execute_ainl_activity() from runtime/wrappers/temporal_wrapper.py, which runs the embedded IR via RuntimeEngine (same contract as LangGraph hybrid).
  2. *_workflow.py — defines a workflow that invokes that activity with workflow.execute_activity, including starter start_to_close_timeout and RetryPolicy you can tune.

Generate artifacts

From the repository root:

python3 scripts/validate_ainl.py path/to/workflow.ainl --emit temporal

Writes <stem>_activities.py and <stem>_workflow.py in the current directory (or use -o — see below).

-o behavior

  • Omitted: files go to the current working directory; base name is the source file stem.
  • Path ending in .py: directory is the parent, base name is the stem (e.g. -o /tmp/job.py → /tmp/job_activities.py).
  • Existing directory: files are written inside it with the source stem as base.
  • Non-existent path without suffix: treated as directory/basename prefix (e.g. -o out/monitoring → out/monitoring_activities.py if out/ exists or is created).

Dependencies

  • AINL runtime: repo root on PYTHONPATH (the emitted activities module walks up to find runtime/ and adapters/).
  • temporalio: required on the worker that runs activities, and for importing the generated workflow module (pip install temporalio). The activities file registers @activity.defn only when temporalio is installed; otherwise run_ainl_core_activity_impl remains callable for local smoke tests.

Input / state contract

The activity accepts a single dict payload. Keys are merged into the AINL frame for the entry label (e.g. metric_value, threshold). Reserved optional keys:

  • _label — override entry label id.
  • _strict — reserved for future use (compile path uses strict=True in the wrapper for source-based runs; emitted IR does not recompile).

The activity returns {"ok", "result", "error"} from execute_ainl_activity.

Benefits

  • Compact authoring: keep branching and R adapter steps in AINL instead of growing workflow code.
  • Durability: Temporal handles retries and timeouts around the whole AINL run as one activity (or split into more activities later).
  • Separation: LLM / human steps can live in other activities; AINL remains the deterministic core.

Example

See examples/hybrid/temporal_durable_ainl/ for source .ainl, emitted *_activities.py / *_workflow.py, and regen instructions.

Related