April 2026 Summaries

Harvey AI's article on their internal platform, Spectre, highlights the challenges and solutions associated with using AI coding agents in a team setting. Initially, Harvey's engineers utilized local coding agents like Cursor and Claude Code, which worked well for individual tasks but failed to provide team-wide visibility and collaboration, leading to isolated code development. Spectre addresses this by executing agent tasks in cloud-based sandboxes with secure boundaries, ensuring durable session records for audit trails, and integrating collaboration tools, allowing team members to interact with and guide agent activities. This approach reflects a broader industry trend where companies like Stripe and Lightsprint are developing similar cloud-based infrastructures for scalable and collaborative AI agent deployments, suggesting that organizations relying on AI agents will need to adopt such systems to overcome coordination challenges.

Developers are increasingly moving AI coding agents from local machines to the cloud due to the limitations and risks of running them locally. Local agents are invisible to teams, insecure due to broad access to sensitive data, and limited in scalability as they can only run a few agents at once. In contrast, cloud-based agents offer enhanced visibility, allowing the entire team to observe and understand the agents' actions and decisions, improved security through sandboxed environments that limit access to only necessary resources, and greater scalability by using elastic cloud computing to run multiple agents simultaneously. The adoption of cloud agents also integrates with the use of preview environments, which allow changes to be deployed to live URLs for thorough review in a working application context. This shift facilitates a more informed and efficient code review process, enabling teams to maximize productivity and security while maintaining local development for tasks requiring direct human interaction.

Coding agents like Cursor, Claude Code, and Codex have significantly accelerated individual coding tasks but have not addressed the primary bottleneck in software development: team coordination. The main hurdle is not the speed of coding but the process of idea execution, which involves multiple steps, from planning and prioritization to feedback and revisions, often taking weeks. The proposed solution is a multiplayer development model where the entire team, including product managers (PMs), engineers, and designers, collaborates seamlessly. In this model, PMs articulate desired outcomes rather than technical specifications, AI agents autonomously interpret and implement these, while engineers focus on reviewing for quality and design. This approach would transform the development process, enabling live previews and real-time feedback, thereby significantly reducing the time from concept to deployment. The key to this shift is creating a coordination layer that integrates AI productivity into team dynamics, promising a tenfold increase in overall product development efficiency and output.

The traditional software development lifecycle (SDLC) is outdated, as it was designed for a time when writing code was the primary bottleneck, leading to a lengthy process involving multiple handoffs and significant waiting periods. With the advent of AI coding agents, the engineering process has become significantly faster, yet the surrounding processes, such as planning and quality assurance, remain unchanged, creating new bottlenecks. A proposed new SDLC streamlines this by collapsing traditional roles into a unified workflow that focuses on four main steps: describing, planning visually, previewing live, and shipping. This approach drastically reduces cycle times from weeks to hours by minimizing handoffs and waiting periods, allowing teams to focus on more strategic decisions and enabling them to ship products much faster. Transitioning to this model requires gradual adoption, starting with smaller features and expanding as trust in the process develops, ultimately leading to increased productivity without increasing engineering capacity.