Execute Python in this session's sandbox. Import tools or a specific module in every call (database, schema, checkpoint, materialized_tables, saved_query, filters, location, municode, text_documents, files, map_image, boundary, spatial_relations, branding) — the python/<module> guides carry the exact signatures.

Requires a viewport — call navigate first. The sandbox can reach any host on the internet. Each execution starts with a fresh namespace: imports and variables do not carry into the next call, but files under /home/daytona persist for the session. stdout and stderr both come back (print() output and tracebacks); the last expression's value is not returned. Tracebacks cite your code as File "<agent-code>" with correct line numbers; frames labeled "<string>" are runtime wrapper, not your code.

Calls that finish within ~45 seconds return synchronously; longer calls are promoted to a background job automatically — you get the job id, the code keeps running (up to 15 minutes), and get_result returns the output (kept for one hour). The wait covers everything in the call, not each query separately. Set background:true to skip the wait for known-heavy work (map_image renders, broad scans). Executions run one at a time, so another execute_python call will queue. If a call still fails with a transport error, the code may have run: check list_outputs before re-running anything that creates artifacts. Only checkpoints can be deleted (delete_checkpoint); tables, saved queries, filters and files cannot.

Do a whole step in one call (query, save the checkpoints, build the map): hosts cap how many tool calls a reply may make. When a call fails, fix it from what it says rather than rerunning it or stopping, and split a query that timed out. Never select bare geometry columns or print whole tables. Something is saved only once its create call has returned it; tell the user plainly when one failed.