"My attention is the bottleneck" | Peter Steinberger, OpenClaw

Agent engineering workflows have evolved from being constrained by token limits and CPU power to the current bottleneck of human attention. Optimizing these workflows requires minimizing the frequency of human intervention by expanding agent capabilities from initial prompting through to verification. Three key technical implementations facilitate this: requiring sanitized transcripts for pull requests to provide necessary context, automating code reviews to ensure design decisions are understood, and utilizing containerized environments like CrabBox to run tests on fresh systems. These tools allow agents to handle complex tasks, such as cross-platform testing and visual verification, while maintaining high-quality outputs. Despite these advancements, human judgment remains essential for high-level architectural decisions, as agents still struggle to grasp the broader context of complex systems.