Code Execution

The Code Execution tool enables AI agents to write and execute sandboxed Python code on the fly to perform exact calculations, data analysis, and algorithmic transformations.

Tool Identifier: code-execution

Why Code Execution?

Large language models excel at natural language understanding and conceptual reasoning, but standard next-token prediction can struggle with multi-digit arithmetic, combinatorial math, or statistical aggregations.

With Code Execution:

  • Deterministic Accuracy: Complex arithmetic (e.g. compound interest, mortgage amortizations, currency conversions) is executed directly by Python rather than estimated probabilistically.
  • Data Transformation: Agents can parse large text tables, sort lists, compute percentiles, and generate formatted summaries.
  • Zero Calculation Hallucinations: When exact numerical results matter, the model delegates the calculation to code and reports the verified output.

How It Works

When an agent needs to perform an analytical or mathematical task:

sequenceDiagram
    autonumber
    actor User
    participant Model as AI Model
    participant Sandbox as Python Sandbox

    User->>Model: "What is monthly payment on a $450k 30-year fixed loan at 6.25%?"
    Note over Model: Generates exact Python script
    Model->>Sandbox: Execute calculation script
    Sandbox-->>Model: Output: "2770.83"
    Model-->>User: "Your monthly principal and interest payment will be 
          

            

              AI Agents & Widgets for AI Builders
            

            

              Build, test, and embed custom AI agents and chat components with AIsh.li.
            

          

        ,770.83."

How to Enable

In the UI

When configuring an agent in the Agent Builder or editing an existing agent, check the box for: ☑️ Code Execution 💻

In Agent YAML

Add code-execution to the tools array:

tools:
  - code-execution

Security & Sandboxing

Code execution occurs inside ephemeral, strictly isolated sandboxes with:

  • No network access from within the script sandbox.
  • Restricted execution time and memory limits.
  • No access to underlying filesystem, host environments, or database credentials.