March 2026 Summaries

Over the past year, the capability of coding agents has significantly improved, surpassing even that of advanced prompting techniques once required. Previously, effective prompts needed detailed instructions, but now they primarily require goals, context, and some preferences, as large language models (LLMs) have become better at following explicit directives. However, it's challenging to predetermine effective rules in the unpredictable real world, necessitating agents to exercise on-the-fly judgment, demonstrating that delegation is more effective than micromanagement. Consequently, there's a shift towards simplifying system prompts, skills, and agent instructions to avoid unnecessary complexity. The author illustrates this with a personal example of a simplified CLAUDE.md file and emphasizes the importance of conveying intent and nuance over rigid orders when writing prompts. Despite improvements, some limitations remain, such as the agents' temporal understanding, highlighting an ongoing evolution in their functionality. Additionally, agents now frequently engage in activities like writing skills, invoking subagents, and orchestrating scripts, with advanced coding agents like Shelley using databases to maintain context and fidelity across interactions.

Organizations are currently navigating the complexities of integrating agentic tools within their workflows, resulting in a diverse range of software production methods. The current climate discourages imposing a single solution and instead promotes experimentation, allowing individuals the freedom to learn and adapt. Central to this approach is the availability of readily accessible, high-performing virtual machines (VMs) that facilitate secure sharing, integration, and data management. This flexibility has led to the proliferation of various agents, automations, and workflows, each evolving to potentially become foundational to software operations. The emphasis is on recognizing the uniqueness of each workflow, as demonstrated by a team of seven individuals utilizing nine distinct workflows, each employing VMs for isolation, experimentation, and collaboration.

In February, the team behind exe.dev made significant updates based on user feedback, introducing features like idea templates for apps such as Gitea and Minecraft, which provide one-click setups with tailored prompts for automatic configuration. They also launched an HTTPS API, allowing users to programmatically manage virtual machines without needing SSH access, and introduced an innovative approach to API tokens using SSH key signing. The AI assistant Shelley received enhancements, including conversation distillation using a language model to maintain context, improved browser tools, and default Markdown rendering in conversations. Shelley also supports internationalization and allows users to purchase LLM credits on demand. Users can keep up with updates through notifications and engage with the team on Discord.

Exe.dev offers a unified command-line interface and API for creating virtual machines, utilizing a Unix-like convention that is familiar to its users. A unique feature is the ability to mint bearer tokens using SSH keys, allowing users to set restrictions and generate tokens without direct interaction with the service provider. The process involves creating a permissions JSON and using an SSH key to sign tokens, which are encoded with URL-safe base64. These self-minted tokens can be used for executing commands like "whoami" and provide functionalities such as scopes, expiry, and revocation by associating them with specific SSH keys. Additionally, exe.dev's VMs include an authentication proxy, allowing users to create bearer tokens scoped to individual VMs for secure access without the need for manual login or cookie manipulation. This system provides a flexible and secure way to manage API access, with more details available in their comprehensive documentation.

Exe.dev has chosen to use persistent virtual machines (VMs) with persistent disks, diverging from the modern trend of stateless or serverless infrastructure, to create a more familiar and user-friendly environment akin to a personal laptop. This approach avoids the complexities and burdens of distributed systems and container registries, allowing users to perform tasks like cron jobs and systemd timers without interruption. By maintaining running VMs, exe.dev aims to simplify development processes and avoid forcing users into complex tool ecosystems or external integrations. This strategy also supports larger projects, offering a scalable infrastructure where users can work seamlessly on existing projects. The company itself utilizes its own VMs to develop, host, and manage various internal tools and services, demonstrating the practical application and effectiveness of their chosen infrastructure model.