Agent Scratchpad

1. Concept Overview & Systemic Problem

When solving a complex mathematical equation or planning a complicated travel route, you rarely write the final version immediately:

• You take a scratchpad (a piece of paper).
• You make intermediate calculations, cross out incorrect options, and draw arrows.
• Only when the answer aligns do you transfer the final result to the form.

Artificial intelligence requires the same kind of scratchpad.

Agent Scratchpad is an internal field for the model's monologue:

• Here, the agent keeps in mind: “What have I already done? What is my next step? Why did the previous request return an empty array?”.
• As soon as the scratchpad is filled with the correct conclusion, the agent performs the action or formulates the final response to the user.

A practical analogy: a private thinking zone where the AI is allowed to doubt, make mistakes, and analyze its steps.

2. Architectural Taxonomy & Mental Model

┌─────────────────────────────────────────────────────────────┐
│                 AGENT'S CONTEXT WINDOW                     │
├─────────────────────────────────────────────────────────────┤
│ 👤 USER REQUEST: “Update client status #12 to VIP”        │
├─────────────────────────────────────────────────────────────┤
│ 🧠 HIDDEN SCRATCHPAD (<scratchpad>):                       │
│    “Yes, I need to find client 12. First, I will check    │
│    if such an ID exists in the system using the get_client  │
│    function. Calling get_client(12)...                     │
│    I see the result: the client's total purchases are 150,000 UAH. │
│    According to the regulations, this corresponds to VIP status. │
│    Now I can safely call set_status('VIP').”                │
├─────────────────────────────────────────────────────────────┤
│ 💬 USER-FACING OUTPUT:                                     │
│    “Client status #12 successfully updated to VIP!”        │
└─────────────────────────────────────────────────────────────┘

3. Technical Pipeline & Internal Mechanics

1. Protection Against Impulsive Actions: Without a scratchpad, the model attempts to invoke a function at the first word and often substitutes fabricated or incomplete parameters.
2. Action History Control: In the scratchpad, the agent records a list of already checked sites to avoid going in circles with the same links.
3. Debugging Convenience for Developers: If the agent behaves oddly, the engineer opens the scratchpad logs and sees verbatim which line of reasoning led the model to the erroneous conclusion.

4. Production Engineering Scenarios

01. Enhancing Decision-Making Accuracy

Incorporating a scratchpad significantly improves the accuracy of decision-making by allowing the agent to evaluate multiple hypotheses before executing commands.

02. Streamlining Debugging Processes

Developers can quickly identify issues by reviewing the scratchpad logs, which detail the agent's thought process leading to unexpected behavior.

03. Reducing Error Rates

Implementing a scratchpad in the system prompt can reduce error rates by 40-50%, as agents are required to articulate their reasoning before taking action.

5. Pitfalls, Common Mistakes & Security

When creating your own agent, always include a requirement in the system prompt to maintain a scratchpad: “Before invoking any tool, be sure to write your thoughts and plan in the <thought>...</thought> block.” This practice is crucial for minimizing errors and ensuring the agent's reliability.