How agent2model compares

The first question people ask is "isn't this just DSPy / LangGraph / distillation?" Short answer: no — they optimize or orchestrate a procedure at runtime; agent2model removes the runtime machinery by baking the procedure into a small model's weights. This page lays out the differences honestly, including where the other tools are the better choice.

The one-paragraph version

If your agent procedure is stable (a support flow, a booking flow, an onboarding script), agent2model compiles it — from a LangGraph StateGraph or a YAML flowchart — into the weights of a small open model (Qwen 3B/8B), so at runtime there's no orchestrator and no per-turn frontier call. The win is cost, latency, and privacy: a self-contained model you own and run offline. Prompt-optimizers (DSPy/GEPA) make a frontier model follow a procedure better but keep a runtime program; orchestrators (LangGraph/CrewAI) run the procedure live every turn. agent2model is the only one that deletes the runtime layer.

At a glance

	agent2model	DSPy / GEPA	LangGraph / CrewAI	Generic distillation
What it optimizes	the model's weights	prompts / programs	nothing — it runs the program	the model's weights
Runtime orchestrator	none	yes (the compiled program)	yes (every turn)	n/a
Per-turn frontier calls	none	yes	yes	n/a
Input	LangGraph graph or YAML flowchart	a DSPy program + metric	your graph/crew	a teacher + dataset
Runs offline / private	yes	no (calls a provider)	usually no	yes
Target	one stable, multi-turn procedure	any task with a metric	any agent app	general capability
Ships procedure-adherence eval	yes	no (task metrics)	no	no

Tool by tool

vs DSPy / GEPA

DSPy and GEPA compile too — but they compile to prompts and program structure, found by optimization/reflection, and the LM + program stay in the loop at runtime. They shine when you want a frontier model to perform a task better and you're fine paying per call. agent2model compiles to weights and drops the runtime program entirely. They're complementary: you could use DSPy-style optimization to design a strong procedure, then agent2model to bake the stable result into a cheap local model.

vs LangGraph / CrewAI

These are runtime orchestrators — they execute the procedure live, re-sending it to the model every turn. That's exactly the cost agent2model removes. agent2model is additive, not a replacement: you can point compile straight at an existing LangGraph StateGraph and it extracts the nodes/edges/decisions for you. Keep LangGraph for procedures that change often or need live tool-use; compile the stable ones into weights. (Tool-use during compiled inference is on the v2 roadmap, not v1.)

vs generic distillation / fine-tuning

Classic distillation clones broad capability from a teacher. agent2model is narrower and that's the point: it internalizes one declared, multi-turn conversational procedure, generating synthetic conversations by traversing the flowchart so the model learns the behavior — and the flowchart never appears in the training data. It deliberately uses full-parameter SFT, not LoRA (the source paper's companion shows LoRA fails to internalize procedures at any rank), and ships an eval harness to prove the procedure stuck.

When NOT to use agent2model

Be honest with yourself — reach for something else if:

• Your procedure changes frequently — you'd recompile constantly; keep it in LangGraph.
• It needs live tool-use / external API calls mid-conversation — that's v2, not v1.
• It's multi-agent with handoffs — out of scope for v1.
• You have no stable procedure yet — design it first (DSPy can help), then compile.

The honesty caveat

The headline 128–462× cheaper / near-frontier quality figures are the paper's (Dennis et al. 2026) and have not yet been reproduced in this repository — reproduction is in progress; status lives in benchmarks/. And note a nuance: a same-author follow-up shows in-context prompting alone nearly matches frontier quality, so quality parity isn't the pitch — cost, latency, and privacy are. What this repo ships today is the open pipeline plus a procedure-adherence eval harness; the numbers will be filled in as the reproduction lands.