Agentic Workflow is the most common enterprise scenario.
Today I read a sharing about how to define a stable and reliable Agentic Workflow within the enterprise.
To summarize briefly.
A few key definitions: Enhanced LLM: combines large language models (LLM), tools, and memory.
WorkfLow with LLM: Place enhanced LLM on a static, predefined path.
Workflow Agent: A predefined workflow run by an Agent, where the workflow is dominant and static.
Aengtic Workflow: By planning and executing workflows based on goals, context, and feedback, it receives a high-level goal and then proactively plans it, breaking down the task into a series of sub-steps and dynamically organizing and calling tools to complete these steps.
The construction and execution of the entire workflow is dynamic, real-time, and can be adjusted and re-planned based on feedback during execution.
This approach is crucial for enterprises because it allows them to automate complex business processes on a large scale, and at the same time, it can fully utilize and revitalize the massive internal microservices and APIs as "tools" that Agents can call.
Agentic Workflow mainly includes the following components: 1 Blueprint generator: Create a natural language high-level plan based on tool capabilities.
2 Planner: Create a low-level mission plan from a blueprint.
3 Executor: Execute the plan.
4 Connectors: combine the output of different tasks and handle rescheduling or termination.
The importance of designing tools from the Agent's perspective requires providing the tool's purpose, description, input/output conventions, and validation checks.
"Blueprint" is a key innovation.
It effectively reduces the cognitive load of the planner by breaking down complex tasks and allows the system to provide a smaller, more relevant toolset for each sub-task, thereby improving planning accuracy and efficiency.
For example, specific examples of preparation for company earnings calls: 1) Summarize previous earnings calls; 2) Collect the latest financial data; 3) Suggest questions based on hot news; 4) Generate the final report.
Planners and executors then complete these subtasks in sequence, and are eventually assembled into a complete brief by the connector.
A strong evaluation framework is very important, including component evaluation and end-to-end evaluation, code-based LLM evaluation and Human-in-the-loop and other technologies to ensure output quality and reliability.
Scenarios where it may not be applicable: For example, fixed repetitive tasks, situations requiring deterministic results, or low-latency/low-cost environments.
Core points: Aentic Workflow is planned and run by Agent, providing large-scale reliability.
"Planning through sub-goal division" is a key design pattern.
The “Plan and Execute” architecture is at the core.
Tools should be complementary to microservices and architecture should be adaptable.
Workflow effectiveness assessments should be treated as first-class citizens.