agent system

When to use it

Use an agent system when a standalone model call can't handle complex tasks requiring multiple steps and resources. Agent systems integrate the model, inference service, and harness to coordinate processes like "fetch data, analyze, and report results."

Quick example

In Anthropic's infrastructure, an agent system processes tasks that require both inference and decision-making. The model predicts outcomes, the inference service manages execution, and the harness coordinates retries and external tool use. Anthropic's setup is a complete agent system, where each component plays a specific role in task processing.

Ecosystem

Agent systems sit at the core of AI task processing, with each component playing a distinct role.

        ┌─ inference service ─┐
model →│     agent system      │→ task completion
        └─ harness ──────────┘

Misconceptions

Trade-offs

• Integration — requires careful coordination of components
• Flexibility — adds complexity in managing multiple parts
• Scalability — demands robust infrastructure to handle tasks