Quickstart¶

This walks the full compile → generate → train → eval pipeline on the travel-booking example. Data generation and evaluation call the Anthropic API; training and serving need a GPU (use Modal if you don't have one).

Install¶

python -m venv .venv && source .venv/bin/activate
pip install agent2model              # core: compile + generate + eval

Optional extras (install only what you need):

Extra	Adds
`train`	torch, transformers, TRL, DeepSpeed — the fine-tuning stack (GPU)
`serve`	vLLM — the OpenAI-compatible serving endpoint (GPU)
`cloud`	the Modal client for the cloud recipes
`report`	matplotlib — renders the PDF eval report
`openai`	OpenAI-compatible client for the served-model eval condition

pip install "agent2model[train]"     # e.g. for local fine-tuning

Set your Anthropic key (used by generate and eval):

export ANTHROPIC_API_KEY=sk-ant-...

1. Compile¶

Validate a workflow and emit the canonical IR. Works on a YAML flowchart or a .py file defining a LangGraph graph.

agent2model compile examples/travel_booking/flowchart.yaml --out build/travel

This writes build/travel/flowchart.json. If the flowchart violates an invariant (unreachable terminal, dead-end cycle, …) the command prints every problem and exits non-zero. See the IR spec.

2. Generate¶

Generate synthetic training data by walking the flowchart with Claude. The command prints the expected cost before starting and the actual cost after, is resumable (checkpoints to build/travel/generation_state.json), and stops if the --budget cap is hit.

agent2model generate build/travel --n 2000 --model claude-sonnet-4-5 --budget 60

Output: build/travel/dataset.jsonl (HF chat-template) and build/travel/cost.json. Useful flags: --seed, --max-concurrent (default 10).

3. Train¶

Full-parameter fine-tune a base model on the dataset. Pick a --size preset (3b single-GPU, 8b DeepSpeed ZeRO-3); --base and --epochs default to the preset.

agent2model train build/travel --base Qwen/Qwen2.5-3B-Instruct --size 3b --epochs 20

The best checkpoint (by held-out eval loss, default 90/10 split) is saved to build/travel/model/best. LoRA is not supported — --lora is refused. See the Training guide.

4. Evaluate or serve¶

Evaluate the compiled model against baselines on the 5-criterion rubric, with bootstrap CIs, significance tests, failure rates, and cost:

agent2model eval build/travel --baselines in_context,langgraph --n 200

Outputs build/travel/eval_report.json and eval_report.pdf. To include the compiled model as a condition, serve it and pass --served-url. See the Evaluation guide.

Serve the compiled model behind an OpenAI-compatible endpoint:

agent2model serve build/travel --port 8000
# POST http://localhost:8000/v1/chat/completions

pip install "agent2model[cloud]"
agent2model cloud setup   # one-time, idempotent
modal run -m agent2model.cloud.modal_app::reproduce_travel

This runs generate → train → eval on Modal and produces a compiled 3B model within 5% of the paper's Table 1. The full first-time walkthrough — wizard, doctor checklist, cost-confirmation prompt — lives in the Cloud quickstart; the per-function reference is in Cloud deployment.