An agentic workflow organizes an AI agent’s actions into a coordinated process for achieving a goal. Its core component interacts with an LLM, while the workflow determines how work is composed. section 12.0 section 12.4
The workflow builds on an augmented LLM equipped with retrieval, tools, and memory. These capabilities provide context, enable interaction with data and APIs, and preserve historical context. section 12.1
Common workflow patterns include:
- Prompt chaining submits prompts in sequence, using one model output as the next input. section 12.1
- Routing sends a prompt to the specialized task best suited to handle it. section 12.1
- Parallelization divides work into tasks that can execute in parallel. section 12.1
- Orchestrator-workers delegates tasks from an orchestrator LLM to worker LLMs and aggregates their results. section 12.1
- Evaluator-optimizer loops between response generation, evaluation, and feedback. section 12.1
Prompt chaining with Spring AI
Start with prompt chaining as a foundational implementation pattern, then compose it with routing or parallelization to form more advanced workflows. Spring AI’s ChatClient can serve as the LLM-interacting component. section 12.2 section 12.4
