Building Your First AI Agent: From Chatbots to Goal-Oriented Models

Every developer faces the moment: “It's great at answering questions, but why can't it just do the work?”

What is an AI Agent?

Simply put, an AI Agent is a combination of LLM (model) + state + tools + loop.

Key Features

• Operates on a goal-oriented basis rather than merely conversational
• Utilizes a repeating loop of “think → act → observe → next action”
• Designed to call upon tools like calculators, files, databases, and APIs

Difference between Chatbots and Agents

• Chatbot: Input → Text Response
• Agent: Goal → Plan → Execute Tools → Validate Results → Next Action For instance, if you ask, “Organize this week's development tasks,” a chatbot just lists them. An agent, however, would check your calendar/issues tracker, prioritize, and draft a summary document.

Planning Your Agent: Start with Minimal Features

Attaching a huge framework from the beginning makes debugging tough. A minimal agent just needs these four components:

1️⃣ State: Stores goals, logs, and interim results 2️⃣ Policy: Decides the next action (model call) 3️⃣ Tools: Executes real actions (functions) 4️⃣ Loop: Repeats until termination conditions are met

Tool Design Tips

• Tools should focus on clear result-oriented tasks (calculating, fetching, storing) rather than “text generation.”
• Keep input/output simple for enhanced reliability.
• Always account for potential failures (exception handling, retries).

Hands-On Code: Creating a Tool-Based Agent in Python

Here's a simple agent with calculator and memo-saving tools. The model assumes OpenAI API compatibility, focusing on “select tool → execute → observe and react.”

import json

# --- Tools ---
def tool_calc(expression: str) -> str:
    try:
        # Demo purposes: Use a safe parser in real services
        result = eval(expression, {"__builtins__": {}})
        return str(result)
    except Exception as e:
        return f"calc_error: {e}"

MEMO = []

def tool_save_memo(text: str) -> str:
    MEMO.append(text)
    return f"saved({len(MEMO)}): {text[:30]}"

TOOLS = {
    "calc": tool_calc,
    "save_memo": tool_save_memo,
}

# --- Agent loop (assumes model returns 'next action' in JSON) ---
def fake_llm_decide(goal: str, observation: str) -> str:
    """
    In practice, replace with LLM call.
    Example return:
    {"action":"tool","tool":"calc","input":"(12000*3)*0.9"}
    {"action":"tool","tool":"save_memo","input":"Result: ..."}
    {"action":"finish","output":"Final answer ..."}
    """
    if not observation:
        return json.dumps({"action":"tool","tool":"calc","input":"(12000*3)*0.9"}, ensure_ascii=False)
    if "saved" not in observation:
        return json.dumps({"action":"tool","tool":"save_memo","input":f"Calc result: {observation}"}, ensure_ascii=False)
    return json.dumps({"action":"finish","output":"Calculation finished and saved."}, ensure_ascii=False)

def run_agent(goal: str, max_steps: int = 5):
    state = {"goal": goal, "steps": []}
    observation = ""

    for step in range(max_steps):
        decision = json.loads(fake_llm_decide(goal, observation))
        state["steps"].append({"decision": decision, "observation": observation})

        if decision["action"] == "finish":
            return decision["output"], state

        if decision["action"] == "tool":
            tool_name = decision["tool"]
            tool_input = decision["input"]
            observation = TOOLS[tool_name](tool_input)
            continue

        observation = "unknown_action"

    return "max_steps_reached", state

print(run_agent("Apply 10% discount on three items (each $120) and save the results.")[0])

Important Points from This Example

• Separating Decision from Observation simplifies debugging
• Clearly defined termination conditions (step limits, finish action)
• Loop doesn’t break even if tools fail (error strings passed as observations)

Example Scenarios: Agent Ideas for Immediate Use

Combining “repetition + rules + external data” creates immediate value in practical patterns.

• Release Note Agent: Summarizes issue lists → Categorizes changes → Generates release text → Saves files
• Code Review Assistant: Summarizes diffs → Detects risky sections → Creates checklists
• Error Triage: Parses logs → Searches similar issues (internal DB) → Organizes reproduction steps

Common Pitfalls and Solutions

Checklist for Enhancing Stability

• Validate Tool Input: Especially strict with numbers/dates/paths
• Separate Permissions: Differentiating “read tools” and “write tools” minimizes mistakes
• Store Observation Logs: Enables troubleshooting and improvements
• Limit Maximum Steps: Prevents infinite loops

Conclusion: Effective Agents Prioritize Flow Over Dialogue

The essence of creating an AI agent is not just smart models, but designing tools and loops to accomplish tasks. Start with the minimal structure introduced today, and expand further with 🔹API/DB tools 🔹planner stages 🔹verification stages for a robust agent.

#aiagent #development #python #llm #agent #automation #softwaredevelopment #productdevelopment