May 2026 Summaries

The text discusses the importance of writing readable and maintainable TypeScript and React code, especially in the context of AI-assisted development. The author emphasizes that code readability is crucial as it is often read more than it is written, and outlines several patterns and practices aimed at improving code clarity. Key practices include keeping functions and components short, using early returns instead of nested conditionals, avoiding unnecessary curly braces in single-line conditionals, defaulting to const for variable declarations, and modeling variants as discriminated unions. Additionally, the text advises pushing queries and mutations to the component level where they are used and maintaining a consistent use of patterns across the codebase for better scalability and ease of understanding by both human developers and AI agents. The author also highlights the use of tools like the Convex platform to ensure end-to-end consistency in codebases, promoting a seamless integration of database schemas and React components. These practices are intended to ensure the codebase remains scalable, trustworthy, and understandable, even as AI becomes more involved in code generation.

Convex has introduced the first wave of community-built Convex Components, expanding the ecosystem beyond the Convex team by launching a Component Authoring Challenge. This initiative has led to the creation of seven new components designed to address common backend needs like web scraping, debouncing, file storage with CDN, email campaigns, image generation, media processing, and payments, allowing developers to integrate these functionalities without building from scratch. Convex Components act as isolated mini-apps within a larger Convex app, each with its own database and functions, which ensures safety from data conflicts and simplifies upgrades. The new components, including Firecrawl Scrape, Durable Agents, Convex Debouncer, Nano Banana, ConvexFS, Transloadit, and Loops, offer enhanced features like durable caching, transactional guarantees, and real-time updates, making them robust solutions for specific tasks. This development marks a significant step towards a vibrant community ecosystem where external authors can contribute, raising the quality standards and broadening the range of available backend solutions.

Mike Cann discusses the challenges and innovative solutions in building Convex OS, a browser-based React application with real-time synchronization, focusing on the integration of Stage Hand, a library that uses AI for browser automation. He highlights the difficulties of traditional end-to-end testing, which often requires frequent updates and can become a time-consuming task. Stage Hand, leveraging natural language to control browsers without specifying detailed UI elements, addresses these issues by enabling tests that mimic human interactions more naturally. Cann emphasizes the value of AI in automating quality assurance, especially for smaller companies without dedicated QA teams, by providing a more intuitive and less brittle testing process. He also reflects on the development experience, including the use of Convex's reactivity model to manage state across multiple tabs seamlessly, and the schema design challenges he faced in modeling the OS's components in a database. Cann concludes with insights into the future potential of AI in testing and the broader implications of designing state management systems for web applications, encouraging developers to leverage Convex's backend capabilities to simplify multi-tab synchronization and state persistence.

The integration of PostHog with Convex provides two distinct paths for enhancing app analytics and observability: infrastructure-level log streaming and error reporting, and product-level event capturing with the PostHog Convex component. Log streaming, which requires a paid Convex plan, offers insight into backend operations by tracking function execution and errors, while the PostHog component, accessible on a free plan, allows for tracking user interactions and utilizing feature flags. The setup of these integrations, detailed in the guide, is designed to ensure that teams can efficiently capture the data they need without unnecessary complexity, with each path serving different analytical needs. The inclusion of AI and LLM analytics through PostHog further enriches the data insights by capturing traces, latency, token usage, and cost, making it valuable for teams scaling AI features. By understanding the specific needs of their projects, teams can choose the appropriate path or combine both to gain comprehensive insights into both infrastructure and user behavior.

In a personal exploration of AI-driven coding, Mike Cann examines the capabilities of Claude Code, a terminal-based AI coding assistant, in building a full-stack Convex app without using an IDE. The experiment highlights Claude Code's strengths, including its plugin extensibility, autonomous planning mode, and effective tool usage, particularly within Convex's TypeScript-native API, which allows seamless AI interaction due to its consistent and predictable nature. Despite the impressive speed and autonomy of Claude Code, Cann notes significant drawbacks compared to IDE-based tools like Cursor, such as issues with discoverability, mouse interaction, model switching, and cost control. These challenges make it less practical for routine development tasks where visual feedback and cost efficiency are crucial. Cann concludes that while Claude Code excels in agentic workflows for greenfield projects, Cursor remains advantageous for everyday development due to its intuitive interaction model and flexibility. The exploration underscores a shifting landscape in software development, where developers may need to adapt to roles focusing more on architecture and decision-making rather than solely on coding expertise.

Convex offers a platform that allows developers to build asynchronous AI applications using TypeScript, focusing on a reactive database model that enhances backend efficiency and reliability. It addresses challenges in AI workloads, where traditional request-response models often falter due to the asynchronous nature of AI tasks that can extend beyond a single request. Convex facilitates this by using reactive databases that turn queries into live subscriptions, ensuring that client applications automatically receive updates without the need for polling or manual cache invalidation. The platform supports durable functions, enabling long-running processes to journal each step and resume from checkpoints, thereby handling failures and ensuring continuity even across client disconnections or server restarts. This model supports efficient coordination across multi-node environments and ensures that persistent data and synchronization are maintained, minimizing the need for additional glue code and making the development of scalable, reliable AI applications more straightforward.

The text explores the challenges and solutions for conducting end-to-end (E2E) testing in software development, emphasizing the use of AI-driven tools like Stagehand and Convex to address common issues with traditional selector-based tests. Traditional E2E tests often become brittle due to their dependency on specific UI elements, which can change frequently as applications evolve. This tends to make them expensive and time-consuming to maintain. Stagehand, a Playwright-based library, allows developers to use natural language commands interpreted by a language model to perform tests, which helps in aligning the tests more closely with user intent rather than the underlying DOM structure. The integration with Convex, which offers a reactive backend that can be reset for each test run, ensures that tests are repeatable and isolated without affecting the development environment. The setup includes splitting unit and E2E tests using Vitest projects, maintaining separate environments to avoid data pollution, and utilizing a test-mode authentication that bypasses traditional OAuth flows. This approach not only reduces the cost of testing but also enhances the detection of UX issues by observing how the AI agent navigates the application. The article suggests that while AI-driven tests are not yet a complete replacement for traditional scripted tests, they complement them by providing flexible, exploratory coverage and identifying user experience problems that might not be evident through conventional testing methods.

Mike Cann discusses an innovative approach to providing AI agents with a bash terminal and file system without using Docker, VMs, or containers, made possible by Convex Sandbox, an open-source project by Patrick Frenet. This setup leverages Vercel's Just Bash to create a virtual in-memory file system and uses Convex storage for state persistence, offering a lightweight alternative for agentic workflows. The environment allows AI agents to perform shell-like operations through bash commands, facilitating tasks like file creation, modification, and deletion, while enabling real-time sync across multiple sessions. Although Convex Sandbox is not intended to replace container-based solutions entirely, it presents a compelling option for specific workloads that benefit from a shared state model and instant execution without the overhead of traditional infrastructure. Despite some limitations, such as the inability to persist empty directories and a soft cap on file numbers, this approach offers a streamlined way to integrate shell access into agent applications, taking advantage of the agents' pre-existing knowledge of bash commands.

The text explains the challenges of database transaction management, focusing on the trade-offs between optimistic concurrency control (OCC) and pessimistic concurrency control (PCC) using a metaphor of a kingdom with a castle and accountants. It highlights how OCC assumes infrequent conflicts and validates at commit time, leading to potential retries but preventing deadlocks and system-wide slowdowns, while PCC locks records to prevent conflicts but risks deadlocks and cascading failures if a transaction is slow or gets stuck. The importance of understanding record contention is emphasized, as it can affect database performance and lead to contention issues when multiple transactions attempt to access the same data. The text also critiques the default isolation levels in databases like Postgres, which can silently lose data due to loose isolation levels, and recommends potential solutions like introducing staleness through aggregation or using pre-built components to manage hot records in systems like Convex. These strategies help manage contention effectively and ensure database systems remain efficient and reliable under load.

The conversation with James, CTO of Convex, delves into the company's approach to hiring and its stance on AI usage in interviews and the workplace. Convex eschews the use of AI tools during coding interviews, emphasizing that the purpose of the interview is to gauge a candidate's thought process, reasoning, and problem-solving capabilities without additional noise from AI. James argues that interviews are not simulations of the job but rather a distinct activity aimed at understanding how candidates think under pressure. While Convex does not mandate AI usage for employees, it holds them accountable for the quality of their work, allowing them to use any tools that enhance their productivity. The hiring process at Convex focuses on clarity of thought and judgment, avoiding rubrics in favor of detailed debrief sessions to assess candidate fit and continually refine the team's understanding of what "good" looks like. The company values cultural alignment and ownership, emphasizing that early hires play a fundamental role in shaping and sustaining the company's culture.