Repeat work,
off your plate.
If someone opens the same tabs and assembles the same answer every week, you have a place to start. Build an agent around that job, give it the right tools, and keep a way to review its work.

A useful place to start
Find a job you already understand.
These examples use real public APIs in the demo workspace. They show different ways to put an agent to work, with the original run and setup alongside the result.
Keep up with new research.
A scheduled digest searches a paper index, keeps context between runs and asks a second model to check the answer.
See the recorded digest, including what the checker questioned.
Research digestTurn a list into a useful answer.
A finance workflow takes pasted invoices, reads reference exchange rates and calculates the change before a model writes the note.
The calculation and the written explanation have different jobs.
Exchange-rate noteAnswer a recurring question in chat.
The Dependency Check agent looks up a pasted package list in the npm registry and answers on a published chat page.
The person asking does not need to open the agent builder.
Dependency checkFrom a job to a working agent
Start small enough to check.
Pick an output you can judge: a brief, a lookup or a calculation. Test it against work your team already knows.
Tell it what a good result looks like.
Describe the task, what sources to use and what to do when information is missing.
Connect what the job needs.
Use a public API to try the idea, or connect your own service with its description and credentials. The first connection may need your IT team or a developer.
Choose how the work starts.
Let people ask in chat, run the agent yourself, or set a schedule. Inspect the result before relying on the job.
Keep the right checks in place.
Choose its tools, configure spending and approval limits, and add another model to check its answers. Keep reviewing the cases that matter.
For the person building it
One place to shape the job.
Instructions, tools, memory, models and runs stay with the agent. Start with autonomous behavior, or use a workflow when the steps need to be fixed.
Explore the builderFor the people using it
A way to ask and get the result.
Publish a web chat or add a website widget. Choose another supported channel when it fits your team, with the connection and account setup that channel requires.
See the channel setup guidesStart with the job that keeps coming back.
Bring one task your team knows well. Build the agent, try the work and inspect what happened.