Agentic Workflows

Agentic Workflows & Multi-Agent Orchestration

1. Skill Context

Focus: Designing autonomous AI agents capable of reasoning, planning, executing tools, and correcting their own mistakes over long-running tasks. Triggers: ai-agents, agentic-workflows, react, langgraph, autogen, multi-agent, planning.

2. The Evolution of Prompting

Standard LLM interactions rely on Zero-Shot or Few-Shot prompting, where the model generates a final answer immediately. Agentic Workflows wrap the LLM in a control loop (a state machine) that allows it to interact with the external world (via APIs, code execution, or databases) before returning an answer.

3. Core Agent Architectures

A. ReAct (Reason + Act)

The foundational agentic loop. The agent iterates through a strict cycle:

1. Thought: The LLM reasons about what to do next based on the user prompt and current state.
2. Action: The LLM requests to call a specific Tool (e.g., search_web, read_file).
3. Observation: The system executes the tool and feeds the raw result back to the LLM. (The loop repeats until the LLM's "Thought" decides the final answer is reached).

B. Plan-and-Solve (Planner-Executor)

ReAct struggles with massive, multi-step goals because the LLM loses focus or gets stuck in rabbit holes. Plan-and-Solve splits the brain:

• Planner Agent: Looks at the user request and generates a rigid Markdown checklist of steps. (It does not execute tools).
• Executor Agent(s): Takes one step from the checklist, executes it using ReAct, and returns the result.
• Benefit: The Planner maintains the high-level context, ensuring the system doesn't drift.

C. Multi-Agent Orchestration (LangGraph / AutoGen)

Complex enterprise tasks require multiple specialized agents working together.

• Supervisor Pattern: A routing agent (Supervisor) receives the task, decides which sub-agent is best suited (e.g., the Database_Agent or the Frontend_Agent), routes the request, evaluates the response, and then routes to the next agent.
• Hierarchical Teams: Structuring agents like a human company. A Tech_Lead_Agent reviews the code produced by the Coder_Agent. If the code fails tests written by the QA_Agent, the Tech_Lead_Agent sends it back to the Coder_Agent with feedback.

4. Architectural Anti-Patterns

• Infinite Tool Loops: The agent calls read_file("wrong_path.txt"), gets an error, and blindly repeats the exact same action 50 times, burning through API credits. Fix: Implement hard limits (max_iterations) and prompt the agent to explicitly change its strategy on failure.
• Hallucinated Tools: The LLM tries to call a tool that isn't in its JSON schema. Fix: Strict system prompts and rigid function-calling (JSON mode) enforcement.