Concept

What is code-mode execution?

Code-mode execution is an approach where AI agents write code against a curated SDK inside an isolated sandbox instead of making dozens of brittle tool calls. One reviewed program replaces many round-trips — cutting token overhead while every run is permissioned, isolated, and sealed in a tamper-evident audit record.

In practice

• The agent expresses a task as code against a curated SDK — not as a chain of fragile tool invocations.
• The code runs in a hardened, isolated sandbox under least privilege.
• High-risk actions pass through human approval gates before they execute.
• Every run produces a signed receipt: replayable, verifiable, exportable.

Code mode vs. tool calling

Tool calling asks a model to orchestrate many small API calls, each a chance to fail and each consuming context. Code mode hands the model a safe workshop instead: it writes one program against a vetted SDK, the sandbox enforces the boundaries, and the receipt proves what happened.

Common questions

Why is code-mode execution safer than tool calling?

Because the boundary moves from 'which tools exist' to 'what the sandbox permits.' The agent's code runs under least privilege in isolation, permissions are enforced at execution time, and the entire run is sealed in a verifiable audit record.

What runs the sandbox?

Secure Runtime — the trust subsystem of the 95% platform. It runs as a production MCP server, executing agent code in hardened sandboxes with governance built in.