Migrating from LangGraph to AINL
This guide is for teams comparing LangGraph (or similar prompt-centric Python graphs) to AINL: a compiled workflow language with strict validation, deterministic runtime options, and emit to LangGraph, Temporal, FastAPI,
Migrating from LangGraph to AINL
This guide is for teams comparing LangGraph (or similar prompt-centric Python graphs) to AINL: a compiled workflow language with strict validation, deterministic runtime options, and emit to LangGraph, Temporal, FastAPI, Kubernetes, or the native runtime from a single source.
Why move from LangGraph to AINL
| Concern | LangGraph-style (typical) | AINL |
|--------|---------------------------|------|
| Token cost | Orchestration often re-invokes the model per step or per run | Compile once, run many — no recurring orchestration LLM tokens on deterministic paths |
| Determinism | Control flow can be model-dependent | Compiled graph IR — same inputs and adapter results → same control flow |
| Single source of truth | Target-specific rewrites for Temporal, K8s, etc. | Author in .ainl → emit multiple targets; benchmarks in-repo |
| Audit | Hard to reconstruct “why” from chat or ad hoc code paths | JSONL execution tape, policy gates, strict diagnostics |
AINL does not remove your LLM from the system where you still want reasoning — it removes the LLM from the orchestration plane for workflows you compile and validate.
Side-by-side: prompt-style vs AINL graph
| Aspect | Prompt / LangGraph agent loop | AINL graph |
|--------|------------------------------|------------|
| Control flow | Implicit in model + Python | Explicit labels, J, If, Call |
| Validation | Mostly runtime | ainl check --strict before deploy |
| Side effects | Scattered in tools | Declared adapter operations with capability boundaries |
| Cost model | Tokens every orchestration turn | ~0 orchestration tokens after compile on pure deterministic runs |
Step-by-step: simple LangGraph agent → AINL
1. Identify the skeleton
List states/nodes and edges: “read inbox → evaluate threshold → escalate or noop.”
2. Scaffold
pip install ainativelang
ainl init my-worker
3. Write the core in .ainl
Express branches with If and fixed labels; call adapters with R adapter.OP … ->var. See examples/hello.ainl and examples/monitor_escalation.ainl.
4. Hybrid emit (optional)
Keep an outer LangGraph or host process while the core is AINL-compiled — see hybrid examples under examples/hybrid/ and the comparison tables.
5. Validate and ship
ainl check workflow.ainl --strict
ainl run workflow.ainl --trace-jsonl run.jsonl
6. Benchmark honestly
Use committed CI artifacts in BENCHMARK.md and docs/competitive/ on the site — no hand-waved competitor numbers.