Cloud agents
When the deliverable is a file
A chat reply is not a deliverable if the thing you needed was a document somebody can open.
The short answer
AI agents that produce files write actual artifacts rather than pasting text into a reply. A Polaris worker chooses whether work deserves a file, names it, picks the format from PDF, Markdown, CSV, JSON, HTML or plain text, and uploads it to the task like a human would. PDFs are typeset server-side into a real binary. Written deliverables default instead to pages in the shared document tree.
- File formats
- pdf, md, csv, json, html, txt
- Size limit
- 2MB of text per file
- Default for prose
- A page in Docs
Two destinations, chosen on purpose
Written deliverables go to Docs by default. Briefs, reports, plans and research become pages in the org's document tree, with real block structure: headings, paragraphs, bullets, to-dos, quotes, callouts, code and dividers. The worker picks a parent page whose topic fits, or the page files under Unsorted, and the delivery comment references it so a reader can follow the link.
Files are for when a file format is the point. Tabular data as CSV, structured output as JSON, a board-ready document as PDF. The runtime instructs workers that files complement the document rather than replacing it, which is why a research session usually returns a page and a session producing a dataset usually returns a file.
What happens when a worker attaches a file
- 1
The worker decides and names it
Format and filename are the worker's call, made from what the work needs. The name is sanitised to safe characters before anything else happens.
- 2
PDFs get typeset
A PDF is not text renamed. Markdown-ish content is laid out server-side with a coloured header band carrying the organisation name, real heading and bullet styling, monospaced table rows, and a footer line naming the worker, the product and the date.
- 3
It goes to storage
Bytes are uploaded under a path scoped by organisation, then task, then a timestamp and the filename, with the correct content type for the extension.
- 4
A row makes it real
An attachment record stores the name, path, byte size, MIME type, version number and the worker's member id as the uploader. It is the same record a human upload creates.
- 5
The feed notices
A trigger writes an attached event to the activity feed naming the file and its uploader, so a file appearing on a task is visible to the team without anyone announcing it.
- 6
The delivery describes it
When files were produced, the delivery comment is a tight summary of them rather than a repetition of their contents.
What the runtime enforces
The guardrails around file production, stated exactly.
Empty files are rejected
A file with no content returns an error to the model rather than creating an empty attachment.
Two megabytes of text per file
Above that the tool returns an error the worker can react to, usually by splitting the output.
A failed PDF render fails loudly
If typesetting throws, the tool returns the render error rather than silently uploading raw text with a pdf extension.
Documents are capped
A drafted page takes up to a hundred and twenty blocks, each up to four thousand characters, with block kinds validated against a known list and anything unrecognised falling back to a paragraph.
Versions are preserved
Every attachment carries a version number, and a revision stores the previous path, size and author before the new bytes are written.
Related reading
Agents deliver, humans close
One rule holds the whole product together, and it is a rule about who is allowed to say finished.
The audit trail behind every agent session
If you cannot reconstruct what happened three weeks later, you do not have an audit trail. You have a feeling.
Sharing the agent, not the transcript
There are two different things people mean by sharing, and only one of them survives contact with a second person.
Delivery comment
Work arrives where the task already lives, attributed and reviewable, and the task stays open.
Hire an AI technical writer
The documentation debt on your team is not a writing problem, it is a nobody-has-two-free-hours problem, and this is the worker for those two hours.
Questions people ask
+Does the agent decide whether to produce a file, or do I?
The worker decides, from the shape of the work and its instructions. You influence it by naming the artifact you expect in the task title or description, which is more reliable than hoping. Written deliverables default to a document page unless a file format is genuinely the point.
+Are the PDFs real PDFs?
Yes. They are generated as binary PDFs server-side with a typeset layout, not text files with a changed extension. Headings, bullets and table-ish rows each get their own styling, and the document carries a header band and a footer attributing it to the worker.
+Can a worker revise a file it already delivered?
Yes, through a review session. Comment on the file, send the open comments to the worker, and it writes a complete new version while the previous one is preserved with its version number, path, size and author intact.
+Where do the files actually live?
In your project's own storage bucket, under a path scoped by organisation and task, with a database row recording name, size, type, version and uploader. They are ordinary attachments, reachable the same way a human upload is.
+Can it produce a spreadsheet or a slide deck?
CSV and JSON cover structured data, and PDF covers a document meant to be read as laid out. Native spreadsheet and presentation formats are not produced by the runtime today.
Related
Cloud agents that keep working after you close the laptop
For the person whose agent is brilliant, local, single-player, and dead the moment the lid goes down.
Agents deliver, humans close
One rule holds the whole product together, and it is a rule about who is allowed to say finished.
The audit trail behind every agent session
If you cannot reconstruct what happened three weeks later, you do not have an audit trail. You have a feeling.
Sharing the agent, not the transcript
There are two different things people mean by sharing, and only one of them survives contact with a second person.
Queueing work at night and reading it in the morning
The realistic version of overnight work, including the part where there is no scheduler.
Delivery comment
Work arrives where the task already lives, attributed and reviewable, and the task stays open.
Work log
The record that makes an unwatched run reviewable afterwards.
Hire an AI technical writer
The documentation debt on your team is not a writing problem, it is a nobody-has-two-free-hours problem, and this is the worker for those two hours.