README.md

Hybrid examples (AINL as deterministic core)

This folder holds interop demos: keep policy, branching, and adapter calls in AINL, while a host framework handles orchestration, durability, or agent loops.

Why use AINL as the deterministic core?

• Lower token cost — You compile the graph once; RuntimeEngine runs it without re-spending LLM tokens on orchestration for the AINL-authored portion.
• Strict guarantees — Optional strict compile checks (adapter contracts, dataflow) catch mistakes before production.
• Auditability — Graph-shaped IR, trajectory hooks, and explicit R adapter steps are easier to review than ad hoc imperative glue.

Patterns

Pattern	What wraps AINL	Example
LangChain tools	langchain_tool adapter from user code / workers	langchain_tool_demo.ainl
LangGraph	StateGraph with one node calling run_ainl_graph	langgraph_outer_ainl_core/
Temporal	Activity + workflow calling execute_ainl_activity	temporal_durable_ainl/

Quick links

• LangGraph hybrid — langgraph_outer_ainl_core/README.md · docs/hybrid_langgraph.md
• Temporal hybrid — temporal_durable_ainl/README.md · docs/hybrid_temporal.md
• CrewAI / LangChain tools — enable --enable-adapter langchain_tool and see langchain_tool_demo.ainl; adapter reference in docs/reference/ADAPTER_REGISTRY.md

Decision help

See the one-page guide docs/HYBRID_GUIDE.md (pure AINL vs LangGraph vs Temporal).

To include LangGraph and/or Temporal in benchmark minimal_emit slices (without using full_multitarget), add S hybrid langgraph, S hybrid temporal, or both at the top of your .ainl source — see the guide.

Optional Python deps for running emitted LangGraph / Temporal modules: pip install -e ".[interop]" from the repo root — details in docs/PACKAGING_AND_INTEROP.md.