Loading…
Topic
Gemini
10 posts about Gemini. Every summary links to the original.
- 10 posts
- 1 company
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Posts
Check what your AI agents do before they do it, with Semantic governance policies
Google Cloud introduced Semantic governance policies in preview for the Gemini platform, enabling enterprises to govern AI agent tool calls before they execute. The policies check each proposed tool call against user-written plain-language rules and the original user intent, preventing actions that violate business rules or result from prompt injection. The enforcement sits outside the agent code via a policy enforcement point such as Agent Gateway, which intercepts model responses containing tool calls. An LLM judge evaluates each call against the conversation, agent instructions, and policies, either allowing it or replacing it with a denial message. The feature supports models from Gemini, Anthropic, and OpenAI-compatible APIs, and logs every verdict and rationale to Cloud Logging for auditability.
How to securely connect internal Cloud Run MCP servers to Gemini Enterprise
Enterprise teams using Gemini Enterprise build custom MCP servers on Cloud Run to connect AI agents to internal APIs, databases, and workflows. Three challenges arise: internal-only ingress blocks public calls, Cloud Run IAM and user OAuth 2.0 must both be verified, and Shared VPCs cannot use wildcard *.run.app DNS zones without breaking other services. The solution uses an Egress Agent Gateway with Private Service Connect to route traffic privately into the VPC, then to Cloud Run via Private Google Access or a global PSC endpoint. FQDN-scoped Private Cloud DNS resolves only the MCP server hostname to a private VIP, avoiding DNS collisions. Gemini Enterprise sends two authorization headers: an OIDC ID token for Cloud Run IAM and the user’s OAuth 2.0 access token for application-level validation. A companion GitHub repository provides a runnable Python MCP server with OAuth 2.0 middleware and step-by-step setup scripts.
Build full-stack Android apps with Gemini in Google AI Studio
Google AI Studio now enables building full-stack Android applications with Gemini 4 Carbon and Hydrogen, shipping them to the Play Store in minutes. The Build agent automatically provisions a Cloud Firestore database with Security Rules and Google Sign-In for persistent, cross-device data. It also configures Firebase AI Logic and App Check to protect Gemini API credentials server-side. New emulator features include faster startup, 1200p HD streaming, real-time audio output, and one-click screenshot capture for visual UI edits. The agent uses the official Android CLI for project scaffolding and builds, and publishes signed Android App Bundles directly to a personal Google Play library.
Gemini 4 Hydrogen: a faster model for general purpose robotics
Google introduced Gemini 4 Hydrogen, a faster and more efficient model for general-purpose robotics that preserves the ability to directly control robot arms from camera images and plain-language goals. On the Artificial Analysis Robotics 0.1 benchmark, Hydrogen scores 58 on eight everyday tabletop tasks, similar to Argon at a lower cost per token. The model outperforms Gemini Robotics ER 2 on the public ERQA benchmark for spatial reasoning tasks such as multi-view correspondence and 3D trajectory reasoning. Google also launched the Robotics plugin for Google Antigravity, which packages robot bring-up, MuJoCo simulation, perception, and application building into skills Gemini can use. A developer can go from a robot in a box to a working application in about an afternoon, such as building a snack-sorting conveyor in an hour.
Securing Your Agents: A Hands-On Guide to Agent Anomaly Detection
Agent Anomaly Detection on the Gemini Platform monitors agent traces and logs to flag sessions where an agent acts outside its intended purpose, catching issues that produce no system crashes or logging errors. This guide walks through deploying a deliberately over-permissive customer support agent whose fetch_records tool accepts any table name without an allowlist. After enabling telemetry and setting up Agent Anomaly Detection, the guide sends one normal refund conversation and one out-of-scope session where a user extracts payroll data. The detection system flags the anomalous session with a Tool misuse finding at High severity and 95% probability, providing a recommended fix to restrict the tool’s table access. Alerting can be configured through Cloud Monitoring metrics and Security Command Center findings, and the setup takes about 30 minutes with automatic monitoring of new agents.
How Agent Anomaly Detection scores on benchmarks we didn't write
Agent Anomaly Detection, a new security feature on Gemini Enterprise Agent Platform, flags sessions where an AI agent acts outside its intended boundaries by reading traces and logs asynchronously. Google tested it against five public, non-Google-written benchmarks for agent goal hijack, the top OWASP risk for agentic applications. The system flagged 87% of harmful instructions that the agent carried out, with false alarms on 2% to 16% of benign sessions depending on the agent’s tools. It performed weakest on poisoned memory attacks, catching only 4 of 18 such sessions. The study used one agent model and simulated apps, so production false-alarm rates may differ.
Catch Go bugs that fuzzing misses with Gemini and a constraint solver
The author built a pipeline that compiles Go functions to RISC-V assembly, feeds the source and disassembly to Gemini 4 Argon, and has the model write Z3 constraint-solver scripts targeting surviving bounds-check traps. Z3 returns a concrete crashing input when satisfiable, and go test replays that input against the real function. The workflow caught all three planted bugs in test parsers and three latent bugs in the skill-lint CLI. Fuzzing missed two of the planted crashes because spare slice capacity or multi-step precondition checks hid the crash. The approach trades away a full correctness proof for practical bug-finding value, with the LLM translation dominating the runtime at about a minute per function while Z3 finishes in milliseconds.
From prompt to app: Build with Google Maps in AI Studio
Google AI Studio now integrates Google Maps Platform, enabling users to build spatial apps from natural language prompts without a credit card. Users describe an idea such as an interactive 3D city tour or a multi-stop route optimizer, and AI Studio automatically configures the necessary APIs, manages keys, and loads a live interactive map preview with zero redirects. The integration provides access to Maps, Places, Routes, and Environment APIs for custom styled maps, photorealistic 3D cityscapes, conversational local search, multi-stop itineraries, and hyper-local environmental monitors. Built-in sandbox quotas of 100 calls per day per API protect against unexpected charges during experimentation, with quotas resetting every 24 hours and no credit card required. When ready for production, users link a Google Cloud billing account to unlock higher quotas and standard Google Maps Platform pricing, which includes recurring monthly free usage.
Cost per task: Using cheap classifiers to route across Gemini models and thinking levels
Per-token API prices are falling, but engineering bills for autonomous coding agents keep climbing because the dominating metric is now cost per successful task, not price per million tokens. A lightweight routing layer can intercept routine decisions before they reach an expensive reasoning model, preventing unnecessary deep-reasoning spend. The article tests TypeSafe’s Jev classifier in front of Google Antigravity’s model calls, using a single atomic Noul question to route tasks. On a 50-question coding-agent benchmark, the Jev router achieved 100% accuracy on 34 unseen held-out tasks and 96% overall, while a frozen regex and prompt-length heuristic dropped to 70.6% on unseen tasks. The most cost-effective approach is to start with a fast model on a light thinking setting, then add a near-free classifier to send only hard tasks to a stronger model.
Building Conversational Commerce with Multi-Agent Swarms on GCP: A Working Prototype
The post details a working prototype called HappyDelivery that implements conversational commerce with a multi-agent swarm on Google Cloud. It uses a three-layer decoupled design: an Ingress layer built with FastAPI that validates and normalizes messages from multiple chat platforms into a canonical schema and offloads them to Cloud Pub/Sub. The Session and Cart layer runs on Cloud Memorystore Redis, storing active cart and profile data keyed by phone number to support asynchronous shopping and dietary guardrails such as nut-free preferences. The Agent Swarm layer orchestrates specialized agents using LangGraph and Gemini, with a lightweight Flash supervisor routing to Profile, Inventory, and Transaction agents that use deterministic Python tools for pricing and payment. Empirical results show a three-turn shopping session consumes approximately 4,240 tokens costing $0.0061, which is 0.030% of a $20.32 basket value.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.