From LangGraph to AINL in 15 minutes

If you already think in graphs and tool nodes, AINL should feel familiar — with one shift: the workflow is a compact program that compiles to canonical IR, validates in strict mode, and runs deterministically without the model sitting in the control plane on every invocation.

What you keep

• Explicit nodes, edges, branching, and tool-like R steps.
• A clear entry and exit story (AINL enforces single-exit discipline in strict mode).
• The option to emit LangGraph when you need that ecosystem today (--emit langgraph).

What changes

• You author (or have an LLM author) AINL source, not Python StateGraph boilerplate.
• The compiler checks reachability, adapter contracts, and dataflow — something prompt-looped LangGraph code paths do not give you for free.
• Compile-once / run-many: after compile, execution does not re-spend tokens on orchestration logic.

15-minute path

1. Read Hybrid guide — why AINL + hybrid emit exists.

2. Open examples/hybrid/langgraph_outer_ainl_core/monitoring_escalation.ainl — small deterministic slice.

3. Run strict validate + emit:

   python3 scripts/validate_ainl.py --strict examples/hybrid/langgraph_outer_ainl_core/monitoring_escalation.ainl \
     --emit langgraph -o /tmp/monitoring_langgraph.py
   

4. Compare token size of the .ainl file vs the emitted .py using your usual tokenizer or the repo’s make benchmark / BENCHMARK.md methodology.

5. Keep .ainl in git as the source of truth; treat emitted Python as a build artifact when you need LangGraph deployment.

When to stay on LangGraph-only

• You need only LangGraph runtime features with no interest in multi-target emission or strict compile guarantees.
• Your team never wants a non-Python authoring surface.

Related

• AINL + Temporal — durability without giving up AINL as source.
• Benchmarks hub / Size tables (repo root) — size and runtime evidence.
• Runner discovery for MCP hosts: HTTP GET /capabilities/langgraph on scripts/runtime_runner_service.py.