May 2026 Summaries

The text explores the complexities and inefficiencies in AI workflows, emphasizing the role of prompt caching in reducing costs while potentially obscuring architectural waste. By analyzing nineteen days of Claude Code session data, the author discovered that prompt caching saved 82% of input costs but also masked inefficiencies such as ever-growing prompts, misaligned context blocks, and divergent sub-agent behaviors. Traditional telemetry methods, like logs and billing dashboards, fail to capture the nuanced behaviors of these AI systems, which operate as stateful, branching entities. The author highlights the need for a new form of telemetry focused on session structure and prompt lineage to truly understand and optimize workflow efficiency. This approach can reveal underlying patterns, such as prompt accretion and context dumps, allowing for more informed decisions about maintaining economic and architectural efficiency in AI operations, particularly before potential changes in pricing or system scale disrupt current efficiencies.

The Paper Compute platform, consisting of 25 member repositories across six layers, faced challenges in managing work across multiple repositories, particularly for coding agents like Claude Code, which struggled with tasks spanning different repositories. To address this, the team developed Paper Forest, a system designed to provide a unified view for agents working across multiple repositories without merging them into a monorepo. Paper Forest introduces two key concepts: a "forest," which is a parent repository containing a manifest of all member repositories, and "groves," which are isolated working directories containing git worktrees for each repository. This system allows agents to work as if they were in a single project while maintaining the independent git identity of each repository. Paper Forest also leverages tools like Crumbly for semantic search, making cross-repository navigation and search tractable for agents. This approach improves agent coherence on multi-repo tasks and enhances human ergonomics by simplifying the setup and management of working environments. While it relocates rather than eliminates multi-repo complexity, the model allows agents to carry the system's full picture, thus reducing the burden on engineers. Future developments aim to introduce forest-wide operations and enhance research capabilities through Crumbly.