Trustworthy Agents in Practice

As AI models evolve beyond simple chatbots into AI agents that perform complex tasks such as writing code and managing files, governance is entering a new phase. Claude Code and Claude Cowork are representative examples.

While an agent's autonomy boosts productivity, it also carries the risk of misunderstanding user intent or being exposed to prompt injection attacks. To address this, Anthropic presents five core principles: maintaining human control, aligning with human values, securing interactions, maintaining transparency, and protecting privacy.

Instead of following a fixed script, an agent operates through a self-directed loop in which it repeatedly plans, executes, observes, and adjusts on its own. An agent's behavior is determined by a combination of the following four elements.

• Model: Provides the intelligence and reasoning capability to carry out tasks.
• Harness: The instructions and guardrails the model must follow.
• Tools: Services the model can use, such as email and calendar.
• Environment: The systems the agent runs on and the data access permissions it has.