Why Controls Matter

Agents can be expensive and unpredictable. Without controls, they might:

• Read 1000 files when you expected 10
• Make 100 API calls that cost real money
• Write to the wrong directory and overwrite important data
• Loop forever trying to fix an unfixable problem

Controls are your safety net.

Three Essential Controls

We’ll add three simple but powerful controls:

1. Tool call budget - Limit total number of tool calls
2. Dry-run mode - Simulate actions without side effects
3. Resource tracking - Monitor what the agent does

These patterns scale to production systems.

Control 1: Tool Call Budget

Limit how many times the agent can call tools:

Integrate into agent loop:

def run_agent(goal: str, max_tool_calls: int = 20) -> str:
    budget = AgentBudget(max_tool_calls)
    input_list = [{"role": "user", "content": goal}]
    
    for iteration in range(MAX_ITERS):
        # ... call model ...
        
        for tool_call in message.tool_calls:
            # Check budget before executing
            if not budget.can_call(tool_call.function.name):
                return f"Budget exhausted: {budget.get_stats()}"
            
            # Execute tool
            tool_output = dispatch_tool(name, args)
            budget.record_call(name)
    
    return f"Complete. {budget}"

Control 2: Dry-Run Mode

Test agent behavior without side effects:

Why dry-run matters:

• Test agent logic without making real changes
• Debug tool calls and see what the agent would do
• Validate plans before execution
• Safe experimentation with new prompts or tools

Control 3: Resource Tracking

Monitor what the agent does:

Putting It All Together

Here’s the agent with all three controls:

Comparing Control Strategies

Agent Control Strategies

This animated concept requires JavaScript to be enabled.

Frames:

1. 

   No Controls: Agent can run indefinitely, make unlimited tool calls, and cause unexpected side effects. Risky for production.

2. 

   With Controls: Budget limits tool calls, dry-run mode tests safely, tracking monitors usage. Production-ready.

Quick Knowledge Check

Production Considerations

When deploying agents to production, consider:

1. Cost controls

• Set budgets per user or per session
• Track API costs in real-time
• Alert when costs exceed thresholds

2. Rate limiting

• Limit tool calls per minute
• Prevent abuse or runaway loops
• Protect downstream services

3. Audit logging

• Log all tool calls and results
• Track who approved what
• Enable compliance and debugging

4. Graceful degradation

• Handle budget exhaustion gracefully
• Provide partial results when possible
• Explain what the agent couldn’t complete

Key Takeaways

You’ve added production-ready controls:

1. ✅ Tool call budgets - Prevent unlimited tool calls
2. ✅ Dry-run mode - Test safely without side effects
3. ✅ Resource tracking - Monitor agent behavior
4. ✅ Graceful limits - Handle exhaustion cleanly

These patterns scale to real-world systems.

What’s Next?

In the next page, we’ll cover debugging and testing: how to trace agent behavior, replay runs, and troubleshoot issues.