Run agents · Your App Cloud
Build an AI workflow agent
For work that needs reading and judgment at each step. Give the agent a short list of actions, mark the ones a person approves, and follow its work in a step log.
Built around a purpose
From a buyer's order change to an amendment you approved.
For people who handle requests that someone has to read before anyone can act. That might be a small producer taking order changes from store buyers, or a caterer turning a client's menu changes into kitchen tasks. If one fixed rule covers every case, a rule-based app fits better than an agent.
For illustration: a hot sauce and salsa maker in Oakland, California supplies 30 independent grocery stores on standing weekly orders. Buyers ask for changes in their own words, such as swapping flavors, adding cases or moving a delivery. Bring the product list with case sizes, each store's standing order and delivery days, and the Tuesday noon cutoff for the week's batch. Add 20 past change requests with what you did about each.
An order-change queue showing the store and order each request matched, the proposed amendment, the packing task, the draft reply and the step log. Amendments and replies wait in an approval list, and stopped runs sit in their own list with the step that failed.
This app needs Pro: an agent reading change requests, saved orders and logs, sign-in for buyers and for you, and email replies are backend capabilities. Free publishes frontend-only static apps. Model and service usage is metered separately.
Write down what done means
Tell Studio when a change request counts as done. Done means matched to a store and order, written up as a proposed amendment, turned into a packing task and answered with a draft reply. A hand-off to you with a written reason also counts.
List the actions it may take
Keep the first list short: read the product list and orders, write a proposed amendment, create a packing task, draft a reply to the buyer. Leave off changing prices, offering credits, editing the product list and promising a delivery day the store doesn't have.
Mark the stopping points
Make every amendment and every reply wait for your approval. Then tell Studio where the agent must stop and hand the request to you. Stop at a product it can't match, a store it can't identify, or any mention of damage, a complaint or a price.
Check it against past changes
Run the 20 past requests in the preview and compare each proposed amendment with the change you actually made. Where they differ, adjust the instructions or the product list, not the output, and run that request again.
Read the log, then break a step
Follow one run through its log: what the agent read, what it decided and why, and which action it tried. Then mark the mild salsa as discontinued and send a request that adds two cases of it. The run should stop with a hand-off rather than swap in another flavor.
Scope the first version
What a first version includes.
Change request form
The store, the change in the buyer's words and the delivery it applies to. Staff can paste in a change that arrived by phone or email and note where it came from.
Actions and approval flags
Each action the agent may take, with a flag saying whether it needs your approval. Keep them as records you can edit, and after changing a flag, read the next run's log to confirm it followed the change.
Step log
One entry per step: what the agent read, the store, order and products it matched and why, the action it tried and the outcome.
Approval list
Proposed amendments and draft replies waiting for you, each with approve, edit and reject. A rejection asks for a short reason, which is saved to the log.
Stopped runs
Requests where the agent handed over: the step, the reason, and buttons to run the request again or take it over yourself.
Order history
Each store's standing order with every approved change, who approved it and the reply that went out. When a buyer asks what changed last month, the answer is on one screen.
Test it in the preview
Before you share it.
- Try every action on the list once in the preview, then read the log. Each action should appear with the record it touched, and nothing that isn't on the list should appear at all.
- Point the sample stores at email addresses you own. Before approving anything, check those inboxes: they should be empty. Approve one reply and confirm a single email arrives with the wording you approved.
- Submit a request that mentions a broken case, and another that asks for a better price. Both runs should stop and reach you without a draft reply.
- Ask for a flavor that isn't on the product list. The run should stop and quote the words it couldn't match instead of choosing the nearest product.
- Fix the cause of a stopped run, such as adding the missing product, and retry it. It should finish with one amendment and one packing task, not two of each.
- Signed in as one store's buyer, open another store's order link. If the access you describe for each collection (Pro) limits buyers to their own store, it should not load. Signed out, only the sign-in page should load.
Together in your app cloud
The pieces behind the possibility.
Choose the models and services that support the experience you want to create.
Language models
Reason, write, and understand
Database
Keep data and context for apps and agents
Email
Keep conversations moving
Authentication
Give the right people access
Shape the details
Make it work for your world.
- Check the cutoff with a rule, not the agent. Comparing a request's time with Tuesday noon needs no reading, and a rule gives the same answer on every request.
- An open form would let anyone change any store's order. Have buyers sign in to the app, or keep the form for your staff and paste in what buyers send.
- Start the agent on proposed amendments and draft replies only. Add another action once the log shows the current ones working on real requests, and give each new action an approval flag at first.
- Tell Studio to match products against the product list alone, and keep that list current. If it starts as a spreadsheet, the upload is a snapshot, so make later changes, such as a discontinued flavor, in the app.
A little more context
Good questions.
How does an agent work with an app?
The app holds the change form, product list, standing orders, approval list and log. The agent runs inside that app, reads a request, uses the actions you listed and leaves its work in the log for you to review. Studio's building assistant helps you build the app; it is not the agent that handles orders.
Can it run on a schedule?
On Pro, an app can run agent tasks on a schedule you set, such as working through new change requests each afternoon. You can also add a button that starts a run on a single request. Either way, read the log of the first few runs and confirm every amendment and reply waited in the approval list.
Should this be rules or an agent?
Use rules for anything you can write as one, such as the cutoff or a store's delivery days. Use the agent for the buyer's wording, like a request to swap two cases of green salsa for mild and move it to Friday. This app mixes both: rules check the timing, and the agent reads the wording. If rules cover nearly every case, the guide to an app for repeatable work, linked below, fits better.
Can the agent email buyers itself?
Apps can send email on Pro, so an approved reply can go out from the app. Keep sending behind your approval in the first version. Delivery time and inbox placement aren't guaranteed, so send test messages to addresses you own before a real buyer gets one.
Explore Pricing for plans, included capabilities, and usage details.
Start from what you have
Bring the file you already use.
Keep exploring.
Your ideas, up and running
Your next idea has a home.
Build an app. Run an agent. Make it yours.
Start building