Determine which Agentic AI Pattern to Use

Agentic AI patterns define how your AI agent reasons and makes decisions. An agentic AI pattern is always associated with a single LLM.

For new use cases, first prototype with ReAct so that it can give you a sequence of steps. Once steps stabilize and you understand what runs and in what order, use Plan & Execute for auditable production workflow.

These are the available agentic AI patterns:

ReAct Agentic AI Pattern

Adaptive. Number of AI agents: 1.

The AI agent thinks, takes one action, observes the result, then decides on the next step. The AI agent never has a full plan. The AI agent adapts in real-time based on what tools return.

Use When ...

You have new use cases and you want to prototype You don't know the exact steps in advance For low-stakes tasks The environment is unpredictable Tool outputs are unknown Each step depends on new information You have branching, exploratory tasks such as conversational, searching, browsing, or debugging You need to handle unexpected tool failures creatively Tasks require 1–5 tool calls where each call informs the next Example use cases: Incident triage and root cause analysis Order exception resolution Supplier risk investigation Good to know: If your workflow branches such as "if X do Y, else do Z", use a ReAct agentic AI pattern first. Configure Max iterations in the agentic AI pattern to limit runaway loops. Set Max iterations in the agentic AI pattern to 10–15 for most tasks, 20–25 for deep investigation. Use a maximum of 5–8 tools. Each tool increases reasoning overhead per step. Combine multiple tools into one where possible. Knowledge bases: if the AI agent queries a knowledge base, configure the knowledge base to return source document and chunk reference with each answer.

Plan & Execute

Number of AI agents: 2.

Plan & Execute is carried out by two AI agents that work in sequence. You can use different LLMs for each AI agent.

• The Planner AI agent uses the Plan agentic AI pattern.
• The Acting AI agent uses the ReAct pattern.

The Planner AI agent creates a full plan: an ordered, structured list of tool calls.

The Acting AI agent carries out the plan in the prescribed order, without reasoning about what to do next. The Acting AI agent does not skip steps, reorder steps, or decide to call a different tool.

Use When ...

You have high-stakes tasks You know all steps upfront and they are the same every run Every step must be auditable and repeatable You need an explicit, auditable trace and controlled step sequence Compliance mandates the same checks in every case All data sources are known before action begins Example use cases: New employee onboarding Authorization checking Good to know: The quality of the run is only as good as the quality of the plan Each plan step must name the exact tool, inputs, and expected output. If your workflow has "if X then do Y, else do Z", specify that in the Planner's step definitions or use a ReAct agentic AI pattern instead. Knowledge bases: if the AI agent queries a knowledge base, configure the knowledge base to return source document and chunk reference with each answer.