Building Reliable AI Agents With Claude and Tool Use

Why agents fail in production

Agents that work great in demos often fail in production for three reasons:

1. Unbounded loops — the agent keeps calling tools until it hits a rate limit or timeout 2. Poor error handling — a single tool failure cascades and the agent gives up 3. Unpredictable costs — an agent that "thinks" for 20 steps instead of 3 is 6× more expensive

A production agent needs a budget: a maximum number of steps, a cost ceiling, and graceful degradation when things go wrong.

The minimal reliable agent loop

from openai import OpenAI
import json

client = OpenAI(
    api_key="cf-your-key",
    base_url="https://api.clawfeeder.ai/v1",
)

def run_agent(task: str, tools: list, max_steps: int = 10):
    messages = [{"role": "user", "content": task}]

    for step in range(max_steps):
        response = client.chat.completions.create(
            model="claude-sonnet-4-6",
            messages=messages,
            tools=tools,
        )

        msg = response.choices[0].message
        messages.append(msg)

        # Done — no tool calls
        if not msg.tool_calls:
            return msg.content

        # Execute each tool call
        for call in msg.tool_calls:
            result = execute_tool(call.function.name,
                                  json.loads(call.function.arguments))
            messages.append({
                "role": "tool",
                "tool_call_id": call.id,
                "content": json.dumps(result),
            })

    # Exceeded max steps — ask for a summary
    messages.append({"role": "user", "content":
                     "Summarize what you've found so far."})
    response = client.chat.completions.create(
        model="claude-sonnet-4-6",
        messages=messages,
    )
    return response.choices[0].message.content

Defining good tools

Tools should be narrow and composable. A search_web(query) tool is better than a research_topic(topic) tool that does multiple searches internally.

tools = [
    {
        "type": "function",
        "function": {
            "name": "search_web",
            "description": "Search the web for current information",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {
                        "type": "string",
                        "description": "The search query"
                    }
                },
                "required": ["query"]
            }
        }
    }
]

Cost control

With clawfeeder.ai's flat per-request pricing, you can calculate maximum agent cost before running it:

max_cost = max_steps × credits_per_request
         = 10 × 15  # Claude Sonnet 4.6
         = 150 credits (¥1.50)

You know the worst case before you start. Set max_steps according to your cost budget.

What makes a good agent task

Agents shine for tasks where the path to the answer isn't known upfront: - Research and synthesis (fetch, read, summarize multiple sources) - Code debugging (run, observe error, fix, repeat) - Data enrichment (look up each item, fill in missing fields)

Agents are overkill for tasks with a fixed structure — those are better handled by a single prompt with good instructions.