Treat Gemini Structured Output as a Versioned API to Prevent Agent Breakage

·2 views

When applications rely on structured JSON output from Gemini, that output effectively becomes an API contract, meaning field changes or new enum values can silently break downstream consumers. Developers are advised to embed schema version identifiers directly in persisted output values rather than relying solely on deployment configuration. Google's Gemini SDK supports JSON Schema alongside Zod validation, enabling stricter runtime checks at the application boundary. Structured generation handles syntax and shape validation, but semantic and policy-level checks — such as verifying referenced evidence IDs exist or confirming an action is permitted — remain the application's responsibility. Different validation failures require different handling strategies, and policy violations in particular should be blocked rather than re-prompted until the model produces a compliant answer.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Strong Data Foundations, Not Better Models, Make AI Agents Reliable

Agentic AI systems that let employees query company data in natural language are gaining boardroom attention, but plugging a large language model into a poorly structured data warehouse produces fast, confident, and wrong answers. Runpod's experience building a Slack-based infrastructure query agent showed that improving the underlying data architecture delivered far greater gains than upgrading the model itself. Experts argue that enterprise leaders misdirect budgets by fine-tuning models instead of fixing the data foundation those models depend on. For reliable AI deployment, organizations must prioritize three core pillars: security, data quality, and observability. On the security front, granting AI agents only read-only, role-limited database access is far safer than relying on prompt instructions, which remain vulnerable to malicious injection attacks.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CrowdStrike Identifies Three Tool Poisoning Attack Classes Targeting MCP AI Servers

CrowdStrike has published a taxonomy of three distinct attack classes targeting Model Context Protocol (MCP) servers, the infrastructure layer that AI agents use to select and call tools. Attackers can embed malicious instructions inside tool description fields, manipulating AI agents into exfiltrating secrets or altering behavior without modifying any underlying code. Because the vulnerability exists in natural-language metadata rather than executable code, traditional static analysis and SAST tools fail to detect it. A second attack class, known as shadowing, allows one tool's description to corrupt how an agent constructs parameters for a completely separate tool later in the same session. CrowdStrike's Falcon Guardian, announced this month, is described as the first production-level tool capable of tracing prompts through tool calls to detect such prompt-layer threats.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Browser-Based IT Support Simulator with WebRTC and No Backend

Nairobi-based developer Devine Nyaenya has built OrbitDesk, a browser-based training simulator designed to replicate Modern Workplace IT support shifts covering tools like Entra ID, Intune, Exchange, and Teams. The platform lets users triage support tickets, investigate sign-in and audit logs, and practice client communication — all without real tenant credentials or a backend server. Team voice calls are powered by WebRTC and BroadcastChannel for cross-tab signaling, while Web Speech API handles speech-to-text and text-to-speech interactions. OrbitDesk targets IT support professionals, MSP team leads, and candidates preparing for technical interviews, offering a realistic simulation of SLA-driven, high-pressure support scenarios. The project is built with Next.js 16, React 19, and TypeScript, and is available as a PWA, Electron desktop app, and is packageable for the Play Store and Microsoft Store.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Build a Python AI Assistant on Android Using Termux

Developers can create a functional AI assistant on an Android device using Termux and Python, without needing a computer or complex development environment. The setup requires installing Python and the OpenAI package inside Termux, along with a valid API key from an AI provider. A short Python script handles user input, sends queries to the AI model via the OpenAI API, and prints responses directly in the terminal. Credentials are stored securely in a separate .env file using the python-dotenv package to prevent accidental exposure. The basic assistant runs in a continuous loop and can be extended with features like conversation history or custom commands.

0 comments Read more at DEV Community