Agentic System
A software system driven by large models that can autonomously plan, invoke tools, execute multi-step tasks, and interact with its environment. Typical components include a browser, a code interpreter, and memory modules.
Timeline (last 90 days)
An autonomous loop with no termination condition is unacceptable in production environments.
In the Agent context, a trace refers to recording a task execution as a hierarchical call tree, where the root node is the user goal, child nodes are each model call and tool call, and each node carries input, output, latency, and token consumption.
Coding tasks are particularly well suited for Agent evaluation because the results can be objectively verified by programs, unlike open-ended question answering which relies on subjective judgment.
Building Agents usually does not require a strong machine learning background; the practical difficulties are concentrated at the systems engineering level: tool interface design, state management, evaluation, and monitoring.
A single Agent plus clear tools often solves most problems and is easier to debug than complex multi-Agent collaboration.
All Facts (6)
Anthropic提出了代理式系统（Agentic System）的概念，将其分为工作流（确定性编排）和智能体（非确定性自治）两种架构
90%UnverifiedAn autonomous loop with no termination condition is unacceptable in production environments.
50%UnverifiedIn the Agent context, a trace refers to recording a task execution as a hierarchical call tree, where the root node is the user goal, child nodes are each model call and tool call, and each node carries input, output, latency, and token consumption.
50%UnverifiedCoding tasks are particularly well suited for Agent evaluation because the results can be objectively verified by programs, unlike open-ended question answering which relies on subjective judgment.
50%UnverifiedBuilding Agents usually does not require a strong machine learning background; the practical difficulties are concentrated at the systems engineering level: tool interface design, state management, evaluation, and monitoring.
50%UnverifiedA single Agent plus clear tools often solves most problems and is easier to debug than complex multi-Agent collaboration.
50%