How to Write Reliable Tests for Google ADK TypeScript AI Agents

·1 views

Testing AI agents is notoriously unreliable when assertions target exact model-generated text, since semantically identical responses can vary in wording and cause false failures. Google's Agent Development Kit (ADK) for TypeScript structures agents, tools, and sessions as code primitives, giving developers stable points to enforce deterministic contracts around non-deterministic model behavior. A recommended four-layer testing pyramid places fast, cheap unit tests for tools and adapters at the base, with live-model evaluations reserved for deliberate, small-scale human-reviewed assessments at the top. Tools built with FunctionTool and Zod schemas should first be tested as plain TypeScript functions, completely independent of Gemini or ADK's event loop. For integration testing, the approach involves running the agent via InMemoryRunner, collecting events, and converting them into a concise application-owned summary to avoid tight coupling to internal ADK event structures.

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)

What Backend Frameworks Actually Do: A Deep Dive Into Core Mechanics

A technical explainer published on DEV Community breaks down the common problems that backend frameworks like Express, Django, Rails, Laravel, Spring, Gin, and FastAPI are all designed to solve, regardless of their differing syntax or ecosystems. The article traces the lifecycle of a single HTTP request from the moment it hits a server through routing, middleware processing, business logic, database interaction, and response formatting. It explains that frameworks exist to standardize these recurring patterns so developers avoid rebuilding the same infrastructure on every project. Key internals covered include how routers use exact string matching, pattern matching, or radix-tree structures to map requests to handlers, and how the framework's responsibility begins only after the underlying runtime or server parses the raw HTTP request. The piece argues that understanding these mechanics makes choosing between frameworks a more informed, deliberate decision rather than a matter of language preference alone.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

EU classifies ChatGPT as a search engine, triggering DSA compliance rules for startups

The European Commission has officially classified ChatGPT as a search engine under the Digital Services Act (DSA), extending the law's obligations to any company using OpenAI's API. Businesses must now publish transparency pages, log user queries, appoint a compliance officer, and implement misinformation filters, among other requirements. Non-compliance can result in fines of up to 6% of global revenue or €10 million, whichever is higher. The rules apply regardless of whether ChatGPT is used in a chat or search-style interface, as the EU considers any information-retrieval response subject to DSA oversight. Startups are advised to begin compliance steps immediately, with some deadlines as short as seven days from the ruling.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Semrush Study: ChatGPT and Google AI Mode Show Different Brand Visibility Patterns

A new mini study by Semrush reveals that brand visibility in ChatGPT and Google AI Mode can differ significantly, meaning strong presence on one platform does not guarantee the same on the other. The research compared brand mentions, cited sources, and competitive landscapes across 22 industries, finding that ChatGPT tends to surface a broader range of brands while Google AI Mode may concentrate on a different mix of domains. Semrush's findings are drawn from an analysis of around 126 million real US AI search prompts spanning ChatGPT, Google AI Mode, Google AI Overviews, and Gemini. As AI-generated answers increasingly shape how consumers discover brands and products, the study warns that relying solely on conventional search rankings may overlook a growing discovery channel. Semrush recommends businesses assess each AI platform separately and map how relevant customer questions are answered across different AI environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How React Lazy Loading Improves Web Performance Without Slowing Initial Load

Lazy loading is a technique that defers the loading of images until they are needed, rather than loading all assets when a page first opens. In React applications, this can be implemented natively using the HTML loading attribute, which requires no additional JavaScript libraries. Developers are advised to apply lazy loading selectively — only to below-the-fold images — while ensuring hero images and other critical above-the-fold content load immediately. Image optimization through compression, proper dimensions, and modern formats like WebP or AVIF should complement lazy loading, as the technique does not reduce file size on its own. For more advanced use cases, third-party libraries such as React Smart Image offer additional features including blur placeholders, retry handling, and aspect-ratio support.

0 comments Read more at DEV Community