All Articles
Measure a Managed Agent's Behavior Against Fixed Scenarios Before It Reaches Production
The public-preview Managed Agents run autonomously inside an isolated sandbox, so a small prompt or config change can quietly shift their behavior. Diffing the output once, the way you would for a single prompt, is not enough. Here is how to build a regression harness that runs fixed scenarios repeatedly and judges on pass rate, plus a shadow to canary to full promotion with automatic rollback, all with runnable Python.
When Context Caching Didn't Lower My Gemini Bill — Field Notes on Measuring the Real Hit Rate
When Context Caching is enabled but the Gemini API bill barely drops, this field note measures the real hit rate from usage_metadata, separates TTL churn from fragmentation, and walks through a staged recovery.
Ingested but Never Cited: Pruning a File Search Store with Citation Logs and a Quarantine Window
Most documents in a File Search store are never cited, quietly draining both cost and retrieval quality. Learn to log grounding metadata, surface never-cited documents from real usage data, and prune them safely with a quarantine window — with working code.
Collapsing Video Understanding into One Native Call with Omni Flash
How I replaced an ffmpeg frame-extraction pipeline (7-9 calls per clip) with a single native Omni Flash call, the measured differences, and the boundaries where keeping frame sampling still wins.
Designing Batch Image Costs with Nano Banana 2 Lite: Decide by Measuring
How to fold the fastest, cheapest image model, Nano Banana 2 Lite, into high-volume generation: measuring per-image cost, a two-tier setup with a quality model, and retry handling grounded in real numbers.
Splitting Bulk Image Generation Cost in Two with Nano Banana 2 Lite: A Draft-and-Render Design
A two-tier cost design that routes first-pass generation to Nano Banana 2 Lite and final renders to the standard Nano Banana 2, with a minimal Python router you can adapt.
Building Conversational Translation Into an App: Speech-to-Speech With the Live API
A design walkthrough for adding speech-to-speech conversational translation to an app with Gemini 3.5 Live Translate and the Live API, covering session lifetime, automatic language switching, latency budgets, and streaming cost, with working code.
Catching only the deprecations that touch you — feeding the official changelog to url-context
I found out an image model was being shut down three days before the deadline. Here is a deprecation radar that reads the official changelog through url-context and surfaces only the models I actually use, with working Python and the over-alerting tuning I had to do in production.
When Two Managed Agents Fight Over the Same Repo: External Leases and Fencing for Isolated Sandboxes
Every Managed Agents run gets its own isolated sandbox, so a local lock cannot stop two runs from touching the same repo or record. Here is how I serialize them safely with an external lease and a fencing token.
When Gemini API Leaks Japanese Into Your English Output Once in a While — Field Notes on Measuring the Contamination Rate and Tightening It in Stages
You told Gemini to answer in English, and 3 out of 100 runs slip a Japanese sentence into the tail. Here is why you cannot stop that 'once in a while', and a production pattern that measures the contamination rate as an SLO and tightens it with graded recovery, with working code.
Catch Near-Duplicate Posts Before You Publish — a Topic-Cannibalization Gate with Gemini Embeddings
Once a blog passes a few hundred posts, new articles start cannibalizing old ones in search. This walks through a pre-publish gate that embeds each post's meaning with gemini-embedding-2 and blocks drafts that sit too close to something you already wrote — with runnable code and how to pick the thresholds.
Catching the Rows That Quietly Failed Overnight: A Per-Row Retry Ledger for the Gemini Batch API
A SUCCEEDED batch job is not the same as all-rows-succeeded. From running nightly batches as a solo developer, here is a per-row result ledger, a transient-vs-permanent failure classifier, selective retries, and a guard against retrying permanent failures forever, with a working SQLite state machine.