Assigning exclusive roles, requiring weekly report reviews—how does Anthropic assign tasks to its AI?

According to Dongcha monitoring, Anthropic has recently disclosed internal engineering experience from running a human–machine collaborative team over the past several months. Multiple employees worked together in Slack with multiple agents that had independent system credentials. The agents were directly mounted under the team roster and communication threads, with responsibilities clearly divided and projects autonomously pushed forward, just like human employees.

To help agents effectively integrate into the team, collaboration by default makes all work fully transparent. Because agents rely entirely on retrievable text to understand context, the company sets security boundaries at the workspace level and, by default, opens full access to agents, avoiding tedious single-document authorization decision-making. The team assigns proprietary roles to different agents by writing Skill files (for example, designating a specific agent as a software release manager) to prevent employees from running their own personal AIs from fragmenting team information.

An agent’s autonomy is directly proportional to the reliability it demonstrates. In concrete practice, an engineering manager dispatched an agent to independently fix 500 bugs and required the agent to submit weekly reflection reports that include mistakes and lessons learned to avoid repeating them. To mitigate risks, the team uses a dual-confirmation (Doer-Verifier) mechanism, in which one agent reviews the work of another. When an agent earns sufficient trust and operates independently, the team also trains and guides the agent to learn to save human attention—by batching routine questions, and by setting workload guardrails—to ensure the sustainable operation of the human–machine team.