Running code in an agent

What it does

Some jobs need real computation — crunching numbers, cleaning and de-duplicating data, parsing files, running statistics, or turning raw data into a clear result. You can give an agent the ability to run custom code to do exactly that, without writing a line of it yourself. The agent writes a short program, runs it, reads the result, and reports back to you in plain language.

How it works

When an agent runs code, that code runs inside a secure, isolated sandbox. By design:

• It has no internet access. The code cannot reach any website, API, or database, and it cannot reach Praxivara's systems or anyone else's data. It is completely isolated.
• It works on the data you give it. Because the sandbox has no internet, the agent first gathers whatever it needs with its other tools (web search, a connected app, a file), then hands that data to the code to process.
• It starts fresh every time. Nothing carries over between runs.

Popular libraries are pre-installed — for data work that means spreadsheets, tables, statistics, PDFs and more — so most analysis just works.

Turning it on

Code execution is an optional capability you add to an agent. The easiest ways:

• Start from the Code Runner template in Explore Templates — a ready-made agent built to compute and analyze data.
• Ask the builder when you create or edit an agent, for example “give this agent the ability to run calculations and analyze data.” The builder adds the capability for you.

What it's great for

• Calculations and formulas that go beyond simple math.
• Cleaning, de-duplicating, and reshaping lists and tables.
• Parsing CSV, JSON, XML, or messy text.
• Summary statistics, grouping and totals, and spotting trends.
• Generating a file — a report, a chart, or a spreadsheet — from your data.

Good to know

• The agent brings the data. If you want it to work on something from a website, a spreadsheet, or a connected app, make sure the agent can fetch that first — the code itself can't go and get it.
• Keep data sizes reasonable. Very large datasets should be summarized or sampled first; the sandbox is tuned for focused tasks, not bulk dumps.
• No passwords in code. Because the sandbox has no network, credentials would be useless there and aren't allowed.
• It runs as part of a run. You'll see the work in the agent's Activity, and any results land in Deliveries.