Agentic Workflow

The classic paradigm of Agent workflows is the separation of the planning layer and the execution layer, where the planning layer is responsible for understanding goals, decomposing tasks, and generating execution plans, while the execution layer handles actual operations via dedicated tools or smaller models.

Andrew Ng pointed out in his AI prompting course that skillful use of AI has become one of the most impactful skills

The core challenges facing multi-agent architectures include the rationality of task decomposition, the risk of hallucination contagion between agents, and the rising task failure rate caused by error accumulation in long workflows.

A typical Agentic Workflow contains four design patterns: Reflection, Tool Use, Planning, and Multi-Agent collaboration.

baoyu-skills 代表了 AI 工具从单点能力向 Agentic Workflow（智能体工作流自动化）演进的行业趋势

Typical patterns in agentic workflows include generating a plan before execution, calling external tools, having multiple sub-agents collaborate with divided responsibilities, and revising outputs through self-reflection.

Agentic workflows are used to build customer support agents, write in-depth research reports, process legal documents, and analyze patient input to provide medical diagnosis recommendations.

吴恩达提出了Agentic Workflow概念，将多智能体系统与大语言模型相结合

Using LLMs to directly generate text for training the next generation of models does not work, but inserting an Agentic Workflow in between to generate high-quality data is useful for training new-generation models.

Multi-round tool calls introduce risks of latency accumulation and error propagation, requiring fault-tolerance strategies to be planned in advance during the architecture design phase.

Agentic Workflow refers to a model autonomously decomposing goals as the executing agent, invoking tools, processing intermediate results, and iterating in a loop until the task is complete.

If Agent tools take the automated execution route, they need to solve engineering challenges such as permission management, sandbox security, and task interruptibility.

The main technical bottlenecks of Agentic Workflows are concentrated in the reliability of tool calls, error recovery mechanisms, and context length limitations.

A tiered architecture in which 'high-cost reasoning models handle planning and decision-making while low-cost specialized models handle content production' is the mainstream approach to AI creation cost optimization today.

The marketing of 'one-shot generation' often obscures hidden costs, including subscriptions and compute, carefully engineered prompts, and the author's familiarity with the environment and ecosystem.

A recipe is a design pattern in Agent workflows that solidifies validated complex task processes into parameterizable templates containing placeholders

Andrew Ng used a writing analogy to explain the difference between the agentic approach and traditional invocation: traditional invocation is like writing everything at once with no ability to revise, while the agentic approach allows for drafting, researching, self-reviewing, and repeated revision.

For agentic workflows involving multi-step tool calls, even slight degradation at any single step can cause the final result to collapse, and without step-by-step evaluation it is nearly impossible to pinpoint the problematic stage.

An agentic workflow refers to having a large language model operate in a looping, multi-step manner rather than generating output in a single pass.