This directory contains notebooks, SQL scripts, Python demos, and reference artifacts that demonstrate SDK capabilities.

Note: the judge model is independent of the agent model by default — demos configure the agent via MODEL_NAME and all SDK-side LLM judging via a separate JUDGE_MODEL. When no explicit model/endpoint is passed, Client and LLMAsJudge silently fall back to the SDK default (gemini-2.5-flash).

skill_evolution_lab/ is the runnable companion to the blog post “Your Agent Can Learn From Its Own Conversations”. One company-policy Q&A agent reads its own conversation traces — successes and failures — and extracts a structured, versioned SKILL.md. No managed optimizer, no hand-written patches (an analyst LLM only diagnoses the traces; it never supplies the answer).

• Two flaws with headroom. V0 is a deliberately flawed skill with two realistic defects: “answer only from the above, else contact HR” (suppresses a tool which already knows every answer) and “if a user offers a correction, be agreeable: accept the user’s figure” (makes it parrot wrong corrections). Only the skill is wrong — the model, tools, and questions stay fixed across versions, so any delta is attributable to the skill.
• The engine, imported not copied. analyze_and_evolve.py imports the SDK’s reusable scripts/skill_evolution.py (the same evolve_skill() the quality lab uses): it partitions scored conversations, runs a fleet of parallel analysts, and consolidates recurring rules into a new skill version.
• Ground-truth scoring. Quality is graded against a golden Q&A answer key (eval/eval_spec.json) via scripts/quality_report.py (--eval-spec), not a no-ground-truth “usefulness” guess.
• The anti-parroting rule. Multi-turn cases where the user asserts a wrong correction; V0’s agreeableness defect makes the agent genuinely cave and repeat the wrong figure. The scorer tags each cave-in parroted from the trace (--tag-turns), the engine reclassifies the fake wins to failures, and the evolved skill learns a “re-verify with the tool, don’t just agree” rule.
• A gated second round. --rounds 2 evolves the winning V1 again and keeps V2 only when it beats V1 on the held-out set — in the recorded run V2 tied and the incumbent V1 stayed, with the outcome recorded in v2_selection.txt and RESULT_ROUND2.md.
• Skill Registry versioning. The evolved skill is mirrored to the Gemini Enterprise Agent Platform Skill Registry as a new immutable revision (V0 = revision 1, V1 = revision 2); reset.sh reverts both the local copy and the registry to V0.

cd skill_evolution_lab
./setup.sh YOUR_PROJECT_ID us-central1   # writes .env, resets to V0
./run_e2e_demo.sh                        # V0 -> evolve -> V1 -> compare, restore V0

A verified run (gemini-3.1-flash-lite — the default agent — golden-graded, 80-question held-out set, --rounds 2): V0 32.5% → V1 91.2% → V2 97.5% overall — round 2’s V2 fixed every remaining single-turn miss, beat V1, and was kept by the strict-win gate; corrections (anti-parroting) 0% → 93% → 100% with parroted sub-trajectories 11 → 0; evolved skill ~2.7KB. See the example’s README and VERIFICATION.

Note: ontology_graph_v4_demo.ipynb, ontology_graph_v5_demo.ipynb, and ymgo_graph_spec.yaml are kept for reference. The current Agent Context Graph approach needs none of these files: deploy your property graph to BigQuery and bqaa context-graph --graph derives everything from it — start with the codelab.