Python SDK: create & train agents

Python is the builder's language in the LyboAI ecosystem. Agents run in the Rust runtime (desktop binary or on-device); the lyboai package drives that runtime over the same protocol the mobile bindings use, and adds the authoring/training loop around it.

1. 1

   Install and connect

   pip install -e sdk/python
   cargo build -p lybo-cli

2. 2

   Create an agent in a few lines

   from lyboai import Client
   
   with Client(data_dir="/tmp/agent") as lybo:
       lybo.install_knowledge(doc_id="h", title="Handbook",
           body="Tickets are answered within 24 hours.")
       lybo.register_skill(skill_id="app.ask", name="Ask",
           triggers=["what does the handbook say"],
           keywords=["handbook"], pattern="retrieval")
       s = lybo.start_session()
       print(lybo.run_to_completion(s, "what does the handbook say?"))

3. 3

   Train routing from labelled examples

   Class-discriminative tf-idf picks keywords that separate your skills; the phrases also build the on-device semantic centroids.

   from lyboai import train_triggers
   learned = train_triggers({
     "faq.ask": ["what does the guide say about refunds", "search the faq"],
     "task.plan": ["plan my week", "break this goal into steps"],
   })

4. 4

   Build → sign → evaluate → gate CI

   from lyboai import PackBuilder, sign_pack, evaluate
   
   PackBuilder("acme.faq", "FAQ", "1.0.0", "acme") \
     .add_knowledge("faq", "Store FAQ", "Refunds take five days.") \
     .add_skill("faq.ask", "Ask FAQ", pattern="retrieval",
                triggers=learned["faq.ask"]["triggers"],
                keywords=learned["faq.ask"]["keywords"]) \
     .add_eval_case("s", "routes", "what does the guide say about refunds",
                    expect_skill="faq.ask", expect_contains=["refund"]) \
     .save("pack.json")
   sign_pack("pack.json", SECRET, "dist/faq.lybopack")
   # install into a throwaway runtime and assert report.ok in CI

5. 5

   Export privacy-safe training data

   Routing/outcome metadata from traces (never raw user content) — the input to tiny-classifier distillation pipelines.

   from lyboai import export_routing_dataset
   export_routing_dataset(lybo, "routing.jsonl")