In this live stream, the creator embarks on building “Slopwatch,” a self-hosted coding agent observability platform designed to track and analyze multiple coding agents’ sessions, emphasizing data privacy, organizational transparency, and integration challenges. Utilizing AI-assisted research and domain-driven design, the project focuses on clear architectural decisions, user consent models, and practical tooling to enhance productivity in AI-assisted software development.
In this rare live stream, the creator explores the idea of building a new project focused on coding agent observability, aiming to create a useful and complex tool for both personal use and viewers. After soliciting project ideas from the audience, the concept of a coding agent observability platform gains traction, with the goal of tracking and analyzing coding agent sessions, token usage, success rates, and productivity across teams. The creator emphasizes the need for a system that supports multiple coding agents like Claude code, Pi, OpenCode, and GitHub Copilot CLI, and discusses the challenges of integrating diverse data sources and schemas into a unified platform.
The stream delves into detailed research on how different coding agents expose hooks and JSON logs for session data ingestion, concluding that a hybrid approach is necessary since no single method fits all agents. The creator experiments with AI tools like Claude and Whisper Flow to assist in refining the project concept, gathering insights on user roles, consent models, session management, and data privacy. A key decision is to build a self-hosted, on-premises platform that respects organizational data ownership and offers both local and server components, including a lightweight sidecar process running alongside coding agents to capture session events.
Naming the project proves to be a fun and collaborative moment, with the community suggesting various options before settling on “Slopwatch,” a catchy and fitting name for the observability tool. The creator documents the research findings and architectural decisions in markdown files within a new GitHub repository, emphasizing the importance of clear, shared language (ubiquitous language) to align the development process. Concepts like sessions, turns, forks, and sub-agents are defined carefully to model the complex branching and resuming behaviors of coding agent interactions.
Further architectural discussions cover authentication strategies, debating between a full OpenID Connect integration versus a simpler token-based system for initial versions. The backend is envisioned as a self-hosted server, possibly implemented in Rust or TypeScript, with a PostgreSQL database for persistence. The frontend would provide dashboards and live spectate features, with polling mechanisms considered sufficient for near-real-time updates. The creator also reflects on the challenges of balancing privacy, developer consent, and organizational transparency, ultimately favoring a model where sessions are visible within the organization to maximize data utility.
Throughout the stream, the creator demonstrates a methodical approach to software design, leveraging AI-assisted research and domain-driven design principles to clarify requirements before coding. The session ends with plans to continue refining the domain model, setting up the repository, and eventually building the platform while producing content for viewers. The creator expresses enthusiasm for exploring new technologies like Rust and maintaining a focus on practical, useful tooling that addresses real-world needs in AI-assisted software development.