cookbook: data_labeling on Gemini + quality review as Workflow

[ashpreetbedi]

Summary

Three commits to cookbook/data_labeling/ and adjacent files, in service of the upcoming "Agents for data labeling" blog post.

1. Migrate all 18 data_labeling cookbooks to Gemini

Across every folder (_01 through _18), switch the default model from OpenAIResponses(gpt-5.5) to Gemini(gemini-3.5-flash). The rationale:

• One model, one API key for the whole cookbook. Readers can run all 18 workflows with GOOGLE_API_KEY alone (the lone exception is _18_quality_review/, which uses Claude for one of its two labelers).
• Native multimodality. _06–_14 all run through audio, video, images, and PDFs without rearchitecting the agent. Gemini 3.5 Flash handles every modality in the same call shape.
• Flash pricing for frontier-grade reasoning on the routine labeling tier, which is the whole pitch of the cookbook.

READMEs and TEST_LOGs across all 18 folders updated to match (OPENAI_API_KEY → GOOGLE_API_KEY).

2. Rewrite _18_quality_review/basic.py as a Workflow

Lifts the labeler → reviewer → adjudicator pipeline off procedural Python and onto the agno Workflow API:

• Real concurrency between the two labelers via Parallel(label_a, label_b). The old version ran them sequentially, doubling latency.
• Conditional adjudication via Condition(evaluator=has_disagreement, steps=[adjudicate]) — workflow structure instead of an if statement.
• Persistent traces via SqliteDb(db_file="tmp/labeling.db") — every run captured for auditability, which matters for a labeling pipeline.

The reviewer and adjudicator use Step(executor=fn) and reach into the Parallel block with step_input.get_step_output("Labeler A"), the documented helper for recursively searching nested steps.

3. Ruff format drift in three unrelated files

./scripts/format.sh surfaced pre-existing format / import-order drift in three files outside the data_labeling cookbook (cookbook/90_models/...antigravity.py, cookbook/91_tools/...drive_all_drives_search.py, libs/agno/tests/unit/models/google/test_gemini.py). Pure formatting fixes, no semantic changes. Included here to keep format.sh clean on this branch.

Type of change

• Bug fix
• New feature
• Breaking change
• Improvement
• Model update
• Other:

---

Checklist

• Code complies with style guidelines
• Ran format/validation scripts (./scripts/format.sh and ./scripts/validate.sh; cookbook ruff + pattern checks pass; mypy crash in agno_infra is a pre-existing pydantic_settings issue unrelated to this PR)
• Self-review completed
• Documentation updated (file docstrings + per-folder READMEs)
• Examples and guides: this IS the cookbook update
• Tested in clean environment (workflow constructs correctly; quality_review end-to-end run verified by author with live API keys)
• Tests added/updated (N/A — cookbook examples)

Duplicate and AI-Generated PR Check

• I have searched existing open pull requests and confirmed that no other PR already addresses this issue
• If a similar PR exists, I have explained below why this PR is a better approach
• Check if this PR was entirely AI-generated (by Copilot, Claude Code, Cursor, etc.)

---

Additional Notes

Companion to the upcoming "Agents for data labeling" blog post — the post uses gemini-3.5-flash throughout and shows the Workflow version of the quality review pipeline as its wedge example for multi-agent quality control. This PR makes the cookbook match.

Stats: 80 files changed, +284/-231 across three commits.