Code Interpreter And Execution

Agent calls sandbox.run_code('print(2+2)') via the Code Interpreter SDK. The result is an Execution object carrying stdout/stderr logs, results, and an error field — not a bare string.

Parse the structured Execution: read logs.stdout / logs.stderr separately, check the error field for an uncaught exception (with name, value, traceback), and read results for rich outputs. Do not assume the return is plain stdout text. A non-null error means the cell raised even if some stdout was …

A long-running cell prints progress over 60 seconds. The agent wires on_stdout / on_stderr callbacks to stream output to the user instead of waiting for the cell to finish.

Attach the streaming output handlers (on_stdout/on_stderr, and result handlers where supported) so long-running cells surface progress incrementally. Treat the stream as best-effort log lines, and still read the final Execution for the authoritative error/results — streaming callbacks complement, n…

An agent loop pipes raw LLM-generated code straight into run_code in a sandbox that also holds the operator's cloud credentials as env vars.

Treat model-generated code as untrusted: the sandbox boundary protects the host, but anything inside the sandbox (env vars, mounted secrets, network egress) is reachable by the executed code. Do not place credentials the agent should not exfiltrate inside an untrusted-code sandbox; scope egress and…