Complex, Multi-Step Tasks
The Poly agent can handle tasks that span multiple steps — reading, searching, writing, and organizing — in a single request. You describe the outcome you want; the agent figures out the sequence of actions needed to get there.
What multi-step tasks look like
A multi-step task might look like:
Find all the vendor proposals in the Procurement folder,
read each one, and create a comparison table as a new
document in that folder
To complete this, the agent:
- Searches the Procurement folder for proposal files
- Reads each one
- Extracts the relevant information from each
- Creates a new document with a comparison table
- Saves it to the correct location
You don't have to break this down yourself — just describe what you want at the end.
More examples
Research and synthesis
Read all the documents in the User Research folder and write a
one-page summary of the main themes
Automated reporting
Find all the weekly status update files from this quarter,
summarize the key accomplishments from each, and create
a quarterly summary document
Data extraction
Go through every invoice in the Finance/Invoices folder,
extract the vendor name, amount, and date from each one,
and create a [CSV](/mime-types/text/csv) with that data
Content pipeline
Find the brand guidelines [PDF](/mime-types/application/pdf), then review the 5 most recent
marketing decks in the Marketing folder and tell me which ones
don't follow the guidelines
Organization + renaming
Look at the files in my Downloads folder, figure out what
project each one belongs to, and move them to the right
project subfolders
How the agent handles ambiguity
When the agent encounters a decision it can't make confidently — like which project folder a file belongs to, or how to handle a file that matches multiple criteria — it pauses and asks you. You can unblock it with a clarification and it continues from where it left off.
If you want fewer interruptions:
Use your best judgment and don't ask me unless you're really unsure
Monitoring progress
For long-running tasks, the agent shows progress inline in the chat. You can see what step it's on and what it's currently doing. You can interrupt at any time and ask it to stop or change direction.
The jobs panel at the bottom of the screen also shows active agent operations that involve file I/O.
Limitations
The agent can handle most file-based workflows, but there are things it can't do yet:
- It cannot send emails or notifications on your behalf
- It cannot connect to external services (APIs, databases) unless explicitly integrated
- It cannot edit binary formats like images, audio, or video
- Very large batches (thousands of files) may be slow — for those, consider narrowing the scope or letting it run in the background