Skip to main content
The first thing worth asking for is something that ends in a file, because a file is what you can go and check. That is the whole difference between an agent with a computer and a chat window: what it made is still there when the conversation is over. Fifteen minutes, no code, and three ways of checking the same result.
1

Ask for the file

Open a new chat and say what you want, in your own words, naming where the file should go:
Two things in that sentence are deliberate. It names an absolute path, so there is no doubt where the file lands — /home/daytona is the agent’s home, and the file tools do not expand ~. And it asks for a file and an answer, so you have two things to check against each other.
2

Watch what it does

While it works you get a collapsed chip — Ran /home/daytona/reports/loa… (+1 more). Click it and the turn opens up.
A Splox turn expanded, showing reasoning, a bash tool call and a file read

The same turn expanded: the reasoning, the program it wrote and ran, and the file it read back

There is no loan-calculator tool on this machine, and none was needed. The agent wrote a program and ran it in its shell, then read the file back to check itself. Two tool calls, then the answer:
This is what went into the first tool call. You never write this; it is here so you can see what “it wrote a program” means.
Worth noticing in the reasoning above it: a fixed payment leaves two cents on the table, so it made the last payment $865.25 and the balance lands on zero.
3

Check it from a different chat

Start a new chat — not another message in the old one — and ask:
It reads the file and answers $383.22, adding up the twelve numbers it found there:
The second conversation knows nothing about the first one. The machine is what they share, and that is the point of this check: the file outlived the chat that made it.
4

Take it off the machine

Still in that second chat:
A Splox chat showing a file read and a download link for loan.csv

A second chat reads a file the first one wrote, and hands it back as a link

Behind the chip is one tool call and what it returned:
The link is served from the machine and lasts an hour (expires_in_seconds: 3600). Opened anywhere — no session, no key — it is the file itself, all 422 bytes of it:
5

Look at the machine yourself

The third check needs no agent at all. Open the inspector on the right of the chat: under Sandbox are Open in VS Code and an SSH command for this machine, good for an hour.
That is an editor and a shell on the same disk the agent has been writing to — /home/daytona/reports/loan.csv is there, as an ordinary file. Reading it that way is inspection, not work: you are checking what your agent did, not taking over from it. Revoke access in the same panel ends both at once. See The chat.

Undoing it

Two different things, and it is worth knowing which is which:
  • Delete the chat (chat menu → Delete) and the conversation is gone. /home/daytona/reports/loan.csv is not: deleting a chat never touches the machine.
  • Delete the file by asking — Delete /home/daytona/reports/loan.csv — and it is gone from the disk while the conversation stays.
  • Stop the machine and neither is lost. The machine’s own settings screen says it plainly: a stopped machine keeps its files and comes back where it left off.

What you just saw

The agent writes code instead of filling in a form. No tool was installed for this job. The model’s tool is a shell, and the answer came from a program it wrote in it. Tools explains why the platform is built that way. Files outlive conversations. /home/daytona/reports/loan.csv belongs to the machine, not the chat — that is what made the second and third checks possible. Every turn is inspectable. The reasoning, each call, its arguments, its result and the bill are all kept against the run, and the footer under the answer says which model ran it and how long it took.

Next

Ask for a tool of your own

Something your agent cannot do yet: ask for it, and watch it write the tool and publish a version.

Drive it from your own code

The same kind of run over HTTP: create it, stream its events, read its outputs.