Agentic Workflows
Definition
Agentic workflows are AI system architectures where LLMs autonomously plan, execute multi-step tasks, use tools, and iterate toward goals, moving beyond simple request-response patterns to dynamic, goal-driven automation.
Why It Matters
Agentic workflows represent a paradigm shift in AI application design. Traditional LLM applications follow a request-response pattern: user asks, model answers, done. Agentic workflows break this pattern. The AI takes a goal, plans how to achieve it, executes steps, evaluates results, and iterates until the goal is met.
This unlocks automation that was previously impossible. An agentic workflow can research a topic by searching multiple sources, synthesizing findings, and generating a report, without step-by-step human guidance. A coding agent can implement features by reading requirements, writing code, running tests, and fixing issues autonomously.
For AI engineers, agentic workflows are the frontier of high-value work. Building systems that reliably accomplish complex goals is harder than building chat interfaces, but also more valuable. This is where implementation skills truly pay off.
Implementation Basics
Agentic workflows combine several patterns:
1. Planning The workflow starts by decomposing a high-level goal into actionable steps. This can be explicit (LLM generates a plan) or implicit (workflow structure defines the sequence). Good planning considers dependencies between steps.
2. Tool Execution Agents interact with external systems (APIs, databases, file systems, browsers). Each tool call is a step in the workflow. Results inform subsequent decisions.
3. Reflection and Evaluation Agentic workflows evaluate their own progress. Did the tool call succeed? Does the current output meet quality criteria? Should the plan be revised? This self-correction is what makes agents “agentic.”
4. Iteration Based on evaluation, the workflow loops back: refine the approach, try alternative tools, gather more information. This iteration continues until the goal is achieved or a stopping condition is met.
Key Design Patterns
- ReAct (Reason + Act): Alternate between reasoning about what to do and taking action. The most common agentic pattern.
- Plan-and-Execute: Generate a complete plan upfront, then execute steps. Good when planning benefits from seeing the whole picture.
- Reflexion: After each attempt, reflect on what went wrong and try again with that knowledge. Good for tasks where learning from mistakes helps.
Practical Considerations
- Set maximum iteration limits. Unbounded loops are expensive and can spiral.
- Log every step. Debugging agentic workflows requires visibility into what the agent tried.
- Start with constrained agents. Limit tools and scope before expanding capabilities.
- Evaluate on end-to-end tasks, not just individual steps.
Agentic workflows are powerful but complex. They require robust error handling, clear stopping conditions, and careful cost management. Start simple and add sophistication as you understand failure modes.
Source
Andrew Ng describes agentic workflows as iterative patterns where LLMs plan, reflect, and use tools, achieving dramatically better results than zero-shot approaches.
https://www.deeplearning.ai/the-batch/how-agents-can-improve-llm-performance/