Assigning dedicated roles, requesting weekly retrospective reports, how does Anthropic assign tasks to AI?

According to Dynamic Beating monitoring, Anthropic recently disclosed the engineering experience of running a human-AI collaborative team internally for several months. Multiple employees collaborated with several intelligent agents holding independent system credentials in Slack. The intelligent agents were directly integrated into the team roster and communication threads, dividing tasks clearly like human employees and driving projects autonomously.

To enable the seamless integration of intelligent agents into the team, the collaboration default was full transparency. As the intelligent agents relied entirely on retrievable text to understand the context, the company set security boundaries at the workspace level and fully opened up to the intelligent agents by default to avoid cumbersome single-document authorization decisions. The team allocated distinct roles to different intelligent agents by writing Skill files (e.g., assigning a specific intelligent agent as a software release manager) to prevent individual AI operations by employees from fragmenting team information.

The autonomy of the intelligent agents was directly proportional to their demonstrated reliability. In a specific instance, an engineering manager delegated an intelligent agent to independently fix 500 bugs and required the intelligent agent to submit a reflection report each week containing errors and lessons learned to avoid recurrence. To mitigate risks, the team used a Doer-Verifier mechanism for double-checking, where one intelligent agent validated the work of another intelligent agent. Once an intelligent agent gained full trust and operated independently, the team also trained and guided the intelligent agent to learn to save human attention by consolidating routine inquiries and setting workload guardrails to ensure the sustainable operation of the human-AI team.