agent-workflow-builder（Agent 工作流程构建器）

Agent Workflow Builder

The Agent Workflow Builder skill guides you through designing and implementing multi-agent AI systems that can plan, reason, use tools, and collaborate to accomplish complex tasks. Modern AI applications increasingly rely on agentic architectures where LLMs act as reasoning engines that orchestrate actions rather than just generate text.

This skill covers agent design patterns, tool integration, state management, error handling, and human-in-the-loop workflows. It helps you build robust agent systems that can handle real-world complexity while maintaining safety and controllability.

Whether you are building autonomous assistants, workflow automation, or complex reasoning systems, this skill ensures your agent architecture is well-designed and production-ready.

Core Workflows

Workflow 1: Design Agent Architecture

1. Define the agent's scope:

- What tasks should it handle autonomously? - What requires human approval? - What is explicitly out of scope?

1. Choose architecture pattern: Pattern Description Use When Single Agent One LLM with tools Simple tasks, clear scope Router Agent Classifies and delegates Multiple distinct domains Sequential Chain Agents in order Pipeline processing Hierarchical Manager + worker agents Complex, decomposable tasks Collaborative Peer agents discussing Requires diverse expertise
2. Design tool set:

- What capabilities does the agent need? - How are tools defined and documented? - What are the safety boundaries?

1. Plan state management:

- Conversation history - Task state and progress - External system state

1. Document architecture decisions

Workflow 2: Implement Agent Loop

1. Build core agent loop: class Agent: def __init__(self, llm, tools, system_prompt): self.llm = llm self.tools = {t.name: t for t in tools} self.system_prompt = system_prompt async def run(self, user_input, max_steps=10): messages = [{"role": "system", "content": self.system_prompt}, {"role": "user", "content": user_input}] for step in range(max_steps): response = await self.llm.chat(messages, tools=self.tools) if response.tool_calls: # Execute tools for call in response.tool_calls: result = await self.execute_tool(call) messages.append({"role": "tool", "content": result}) else: # Final response return response.content raise MaxStepsExceeded() async def execute_tool(self, call): tool = self.tools[call.name] return await tool.execute(call.arguments)
2. Implement tools with clear interfaces
3. Add error handling and retries
4. Include logging and observability
5. Test with diverse scenarios

Workflow 3: Build Multi-Agent System

1. Define agent roles: agents = {"planner": Agent(llm=gpt4, tools=[search, create_task], system_prompt="You decompose complex tasks into steps..."), "researcher": Agent(llm=claude, tools=[web_search, read_document], system_prompt="You gather and synthesize information..."), "executor": Agent(llm=gpt4, tools=[code_interpreter, file_system], system_prompt="You execute tasks and produce outputs..."), "reviewer": Agent(llm=claude, tools=[validate, provide_feedback], system_prompt="You review work for quality and correctness...")}
2. Implement orchestration:

- How do agents communicate? - Who decides what runs when? - How is work passed between agents?

1. Manage shared state:

- Task board or work queue - Shared memory or context - Artifact storage

1. Handle failures gracefully
2. Add human checkpoints where needed

Quick Reference

Best Practices

• Start Simple: Single agent with tools before multi-agent

- Prove value with minimal complexity - Add agents only when necessary - Each agent should have clear, distinct responsibility

• Design Tools Carefully: Tools are the agent's hands

- Clear, descriptive names and documentation - Well-defined input/output schemas - Proper error handling and messages - Idempotent operations where possible

• Limit Agent Autonomy: Constrain the blast radius

- Define what agents cannot do - Require approval for high-impact actions - Implement spending/rate limits - Log all actions for audit

• Manage State Explicitly: Don't rely on LLM memory alone

- Persist conversation and task state - Summarize long contexts to fit windows - Track what has been tried/completed

• Fail Gracefully: Agents will encounter errors

- Clear error messages for the agent to reason about - Retry logic with backoff - Fallback strategies - Human escalation paths

• Observe Everything: Debugging agents is hard

- Log all LLM calls and tool invocations - Track reasoning chains and decisions - Measure success rates by task type

Advanced Techniques

ReAct Pattern (Reasoning + Acting)

Structure agent thinking explicitly:

REACT_PROMPT = """
You are an agent that solves tasks step by step.

For each step:
1. Thought: Analyze the current situation and decide what to do
2. Action: Choose a tool and provide arguments
3. Observation: Review the tool result

Continue until you can provide a final answer.

Available tools: {tool_descriptions}

Current task: {task}

Begin:
"""

Planning Agent with Task Decomposition

Break complex tasks into manageable steps:

class PlanningAgent:
    async def solve(self, task):
        # Step 1: Create plan
        plan = await self.create_plan(task)

        # Step 2: Execute each step
        results = []
        for step in plan.steps:
            result = await self.execute_step(step, context=results)
            results.append(result)

            # Replan if needed
            if result.status == "blocked":
                plan = await self.replan(task, results)

        # Step 3: Synthesize final output
        return await self.synthesize(task, results)

Reflection and Self-Correction

Let agents review and improve their work:

async def solve_with_reflection(self, task, max_attempts=3):
    for attempt in range(max_attempts):
        # Generate solution
        solution = await self.generate_solution(task)

        # Self-critique
        critique = await self.critique_solution(task, solution)

        if critique.is_acceptable:
            return solution

        # Improve based on critique
        task = f"{task}\n\nPrevious attempt issues: {critique.issues}"

    return solution  # Return best effort

Human-in-the-Loop Checkpoints

Integrate human approval into workflows:

class HumanApprovalTool:
    async def execute(self, action_description, risk_level):
        if risk_level == "low":
            return {"approved": True, "auto": True}

        # Send to approval queue
        approval_request = await self.create_request(action_description)

        # Wait for human response (with timeout)
        response = await self.wait_for_approval(
            approval_request.id,
            timeout_minutes=30
        )

        return {
            "approved": response.approved,
            "feedback": response.feedback,
            "auto": False
        }

Memory Management

Handle long conversations and context:

class AgentMemory:
    def __init__(self, max_tokens=8000):
        self.max_tokens = max_tokens
        self.messages = []
        self.summaries = []

    def add(self, message):
        self.messages.append(message)

        if self.token_count() > self.max_tokens:
            self.compress()

    def compress(self):
        # Summarize older messages
        old_messages = self.messages[:-5]  # Keep recent
        summary = summarize(old_messages)

        self.summaries.append(summary)
        self.messages = self.messages[-5:]

    def get_context(self):
        return {
            "summaries": self.summaries,
            "recent_messages": self.messages
        }

Common Pitfalls to Avoid

• Building multi-agent systems when a single agent suffices
• Giving agents too much autonomy without safety bounds
• Not handling tool failures and edge cases
• Forgetting that LLMs can hallucinate tool calls
• Infinite loops when agents get stuck
• Not logging enough to debug agent behavior
• Assuming agents will follow instructions perfectly
• Ignoring cost (token usage) in agent loops