April 2026 Summaries

Over a 10-day experiment, a team built a software factory to create a Notion-style note-taking app called Memo, without any human-written production code, by utilizing a network of background agents to automate the entire software development life cycle (SDLC). The experiment demonstrated the capability of these agents to handle tasks such as planning, implementation, review, deployment, and monitoring, with humans only writing specs and overseeing escalations. The project underscored the importance of detailed specifications as the control surface, as it directly influenced the quality of the end product, revealing that insufficient specs led to product iterations and bug fixes. Despite producing a functional product, the factory struggled with design aesthetics, highlighting the need for human taste and guidance. The factory's continuous improvement was facilitated by feedback loops and self-assessment tools, allowing it to evolve and address its own gaps. The initiative emphasized the shift in engineering roles from product building to factory maintenance, and the potential for standardization in software factories, akin to the development of Kubernetes for orchestration engines.

In an exploration of the evolving landscape of platform engineering, the text delves into the concept of "harness engineering," which emphasizes the use of AI agents to automate and standardize software development processes, shifting the traditional role of platform engineers towards managing these agents and designing the systems they operate on. This approach challenges conventional engineering models by suggesting that with the use of agents, even smaller organizations can achieve efficiencies previously reserved for larger teams, despite the agents' tendency to sometimes deviate from expected behaviors. The narrative suggests a return to foundational platform engineering principles as organizations are compelled to relearn lessons from the past to avoid sprawl and inefficiencies. It highlights the importance of empathy in understanding both developers and AI agents to effectively manage software factories, encouraging experimentation with harness engineering to adapt to future challenges. Through practical examples and live demonstrations, the text underscores the transformative potential of AI in redefining the roles and processes within software development.

In a week-long experiment, the team at Ona documented the creation of a software factory capable of autonomously managing the entire software development lifecycle for a product called Memo, a Notion-like note-taking app. Beginning with an empty GitHub repository, the process saw over 130 pull requests merged and 12,202 lines of code produced without human-written code, relying instead on a series of orchestrated automations. These automations handled tasks such as planning, building, reviewing, and verifying code autonomously, with humans only intervening for product specification, automation design, and escalated reviews. The project demonstrated the potential for AI-driven agents to not just write code but oversee the full lifecycle of a product, although it highlighted the continued need for human input in areas like product direction and addressing design flaws not visible through code. The factory's success in rapidly building a functional product underscored the importance of precise input specifications and robust quality controls to ensure both speed and reliability in autonomous software production.

Veto is a security tool that enhances executable blocking by using content-addressable enforcement, which relies on computing a SHA-256 hash of executable content to block processes before they start. This method allows Veto to block executables based on their content rather than their location, effectively preventing renamed or moved files from escaping detection. Despite its efficacy, Veto initially required manual input of executable paths, which posed challenges as AI agents continuously introduce new executables. The improved version of Veto addresses this by utilizing a BPF-based discovery agent that scans containers and automatically hashes and blocks executables based on their names, facilitating real-time monitoring and enforcement without the need for exhaustive manual path specification. This advancement in the discovery layer ensures comprehensive coverage across all container layers, maintaining the robust enforcement model initially established. Veto, now in early access, offers kernel-level security without requiring extensive filesystem management, aiming to support AI agents in secure production environments.

Ona has undergone a significant redesign to enhance collaboration between users and agents, focusing on creating a seamless workspace that prioritizes context retention and minimizes disruptions. By integrating conversations, code reviews, and environment management into a unified interface, Ona allows developers to maintain their workflow without the constant friction of switching tools. The new layout includes features like a bottom drawer for ports and services, inline code review, and a searchable file tree, which help streamline the development process. This redesign not only simplifies interaction with the platform but also opens the door to new capabilities such as integrated terminals, multi-agent coordination, and managed sub-sessions, thus shifting the focus from environment management to problem-solving. With this update, Ona aims to empower developers by providing a calm and intuitive experience that supports efficient and effective task completion.

Lou Bichard details a project focused on building a "software factory" with AI agents at its core, conducted via daily livestreams from April 14 to April 25, 2026. The initiative involves using Ona, a platform designed to run background agents in cloud environments, where the agents execute all coding tasks, while humans provide strategic direction. This approach aims to explore the potential and limitations of AI agents working collaboratively as a fully automated codebase, a concept referred to as a "self-driving codebase." The project is fully open source, with ongoing metrics and updates available on software-factory.dev, and emphasizes continuous learning by updating the system's harness to improve over time. Interested parties can engage with the project through various online channels, including GitHub and social media, to gain insights into AI-driven software development processes.

Implementing an AI-driven system to automatically approve low-risk pull requests (PRs) significantly increased the efficiency of a software engineering team, reducing the median time to first approval from 2 hours 49 minutes to 3.8 minutes and cutting lead time by 74%. This transformation was achieved by allowing an AI agent to review and approve certain PRs, specifically those involving small changes with no sensitive modifications, provided they met predefined criteria. The process involved a governance model where the AI handled reviews, but a human still performed the final merge, ensuring compliance and security. Following the implementation, the team saw a 215% increase in PRs merged and deploys per week, encouraging engineers to produce smaller, focused PRs that are easier to review and ship. The success of the system underscores the importance of automated classification, clear governance, and human oversight in enhancing developer velocity without increasing team size.

As companies like Stripe and Ramp have demonstrated the potential of background agents in automating significant portions of their engineering workflows, the industry is increasingly exploring the build-versus-buy decision for implementing similar systems. Both companies developed bespoke infrastructure to support their agents, leveraging years of standardisation and specialised platform teams to create environments where agents can autonomously write and review code. The complexity lies in providing a complete development environment, ensuring security, enabling connectivity, and allowing automation and parallelism, which most organisations lack the resources to build from scratch. Ona offers a managed platform that provides these capabilities out of the box, addressing common challenges and enabling companies to deploy background agents effectively within their own infrastructure. This allows organisations to focus on customising the system to their specific needs without the overhead of developing and maintaining the foundational infrastructure themselves.