02 · Agentic AI Patterns

Overview of how AI agents are designed, how they reason, and how they act.
This is the architectural core of every development automation system.

What Is an Agent?

An AI agent is a system where an LLM acts as a reasoning engine that can:

1. Perceive — receive context (code, tickets, test reports)
2. Plan — decide what actions to take and in what order
3. Act — invoke tools (APIs, file systems, browsers, compilers)
4. Observe — receive tool results and update its reasoning
5. Repeat — continue until the goal is achieved

The key distinction from a simple LLM call: an agent runs a loop until a stopping condition is met.

graph TD
    A[Goal / Task] --> B[Agent: Plan next action]
    B --> C{Tool call needed?}
    C -->|Yes| D[Execute Tool]
    D --> E[Observe Result]
    E --> B
    C -->|No| F[Return Final Answer]

Agent Architectures

→ Deep Dive: ReAct & Planning Loops
→ Deep Dive: Multi-Agent Systems

The ReAct Loop

ReAct (Reason + Act) is the most widely used single-agent pattern. The model alternates between a Thought (reasoning) and an Action (tool call) until it reaches a Final Answer.

sequenceDiagram
    participant Agent
    participant Tools
    Agent->>Agent: Thought: What service handles this?
    Agent->>Tools: Action: search_codebase("OrderService")
    Tools-->>Agent: Observation: Found in order-service/
    Agent->>Agent: Thought: I should read the relevant class
    Agent->>Tools: Action: read_file("OrderService.java")
    Tools-->>Agent: Observation: File contents
    Agent->>Agent: Thought: Bug is on line 42
    Agent->>Agent: Final Answer: [diff + explanation]

Plan-and-Execute Pattern

For complex, multi-step tasks with clear structure (like a JIRA feature ticket), a planner first decomposes the task, then an executor works through it step by step.

graph LR
    A[Planner LLM] --> B[Step 1: Read ticket]
    A --> C[Step 2: Find service]
    A --> D[Step 3: Implement change]
    A --> E[Step 4: Write tests]
    A --> F[Step 5: Create PR]
    B --> G[Executor Agent]
    C --> G
    D --> G
    E --> G
    F --> G

Tool Use Best Practices

Agents are only as good as the tools available to them. Well-designed tools:

Stopping Conditions

Agents need explicit stopping logic or they can loop indefinitely:

Agent Memory Types