I tried building AI agents and ended up needing an institution (github.com)

A new technical report highlights the need for institutional frameworks alongside AI agents to ensure that AI can perform real work effectively. The author, after operating a human-managed, AI-staffed team, observed that many failures were not due to the AI's capabilities but stemmed from organizational shortcomings—such as unclear ownership of tasks, lack of proper authorization, and inadequate review processes. To address these issues, the report introduces the concept of "institutional control," which focuses on establishing clear roles, accountability, and documentation within AI workflows. This framework emphasizes essential components such as persistent worker identities, defined roles, evidence-backed decision-making, and mechanisms for stopping actions or revisiting decisions. By moving beyond the simplistic view of “did the agent complete the task?” to more complex inquiries like ownership, authorization, and learnings from outcomes, the report asserts that a robust institutional layer is critical for governing AI work effectively. This shift aims to make AI systems not only more transparent and accountable but also better equipped to manage complex tasks that require human oversight, ultimately allowing teams to focus on higher-level strategic work.