Awesome Agents on Google Cloud

Code for “the same agent, three ways”

One agent (“what does Hacker News think about X?”) over the public BigQuery Hacker News dataset, built on three runtimes. Each method is self-contained and takes the question as an optional CLI argument.

pip install -r requirements.txt
export GOOGLE_API_KEY="your-key-from-aistudio.google.com"
gcloud auth application-default login   # BigQuery runs as you, free tier OK
python method1_interactions.py "What do people on HN think about Rust?"

File	What it shows
`method1_interactions.py`	Interactions API: hand-written execute_sql tool, you run the loop, server keeps state
`method2_adk.py`	ADK: the first-party BigQuery toolset, framework runs the loop, agent discovers schema
`method3_antigravity.py`	Antigravity SDK: prepackaged harness, BigQuery via Google’s hosted MCP endpoint
`method3_container.sh`	Runs method 3 in a Linux container (no google-antigravity wheel for Intel macOS)
`_run_with_gcloud_creds.py`	Test-only ADC shim for this machine, not article material

Test status (2026-08-06):

Method 1: VERIFIED end to end on the Gemini API free tier (gemini-3.6-flash); strong answer with real quotes and usernames.
Method 2: VERIFIED on Gemini Enterprise Agent Platform, the platform formerly named Vertex AI; the API and env vars keep the old name (GOOGLE_GENAI_USE_ENTERPRISE=TRUE, location global); real agent-issued SELECTs confirmed in BigQuery job history. Gotchas: needs backoff retries (in the file) because one question is 4-6 model calls; the toolset’s search_catalog tool 403s on public datasets (dataplex.projects.search denied on bigquery-public-data) and the agent routes around it.
Method 3: VERIFIED end to end on GEAP + the hosted BigQuery MCP endpoint (real usernames, 2026-fresh threads in the answer). Gotcha: the MCP tools take a required projectId argument; name the project in the system instructions. Dockerfile.method3 bakes the image (method3-antigravity) for fast starts.

Serving (Cloud Run, us-central1, all IAM-gated): all three smoke-tested and working. hn-agent-interactions uses the API key from Secret Manager (Interactions API is Gemini-API-only, GEAP rejects it) and must run a flash model: pro models have limit 0 on the free tier (the earlier “blocked” state was HN_AGENT_MODEL=gemini-3.1-pro-preview, not the daily reset; switched to gemini-3.6-flash 2026-08-07 and verified).

Quota notes: free tier measured at 5 requests/min AND 20 requests/day per model; agent loops burst through both. GEAP (ex-Vertex) on a billed project has no such caps (transient 429s handled by retries). HN_AGENT_MODEL switches model.