Introducing Agent mode

Until today, asking Nest for a change worked like a fast pair of hands. You described what you wanted, Nest edited the code, and you checked the preview to see if it was right. That loop is quick, but the checking was always yours.

Agent mode changes who does the checking. Nest now plans the change, makes it across every file it touches, runs your app in a real browser to test it, and fixes the errors it finds before it hands anything back. You get a working result, not a draft to debug.

Agent mode is available today on every paid plan, and on the free plan with a monthly allowance.

What Agent mode does

Ask for something larger than a single edit, like “add team accounts so people can invite colleagues,” and Nest works through it in four steps.

1. Plan. Nest writes a short plan: the tables it will add, the pages it will change, the emails it will send. You can read it and edit it before any code is written.

2. Build. It makes the change across the whole project, from the Postgres schema to the server functions to the UI.

3. Test. It boots your app in an isolated sandbox, opens it in a headless browser, signs in as a test user and clicks through the flow it just built.

4. Fix. If a page throws, a type check fails or a button does nothing, Nest reads the error, makes a fix and tests again.

When the loop finishes, you get a summary of what changed and what was tested, with screenshots of the key screens.

In the beta, 87% of agent runs finished with every generated check passing on the first hand-back, and the number of follow-up messages that started with “that broke” fell by more than half.

It asks when it should

An agent that guesses is worse than one that waits. Nest stops and asks when a decision is yours to make rather than a technical detail. The most common cases are:

• Prices and money. What a plan costs, whether there is a trial, which currency to charge in.

• Permissions. Who can see, edit or delete a record. Nest will not quietly decide that every member can remove a project.

• Design choices. When there is more than one reasonable layout, it shows you the options instead of picking one.

Questions arrive as a short message in the chat, usually with a suggested answer you can accept in one click. If you would rather not be asked, say so in the plan and Nest will go with its suggestion and note the choice in the summary.

You stay in control

Edit the plan first

The plan is a real step, not a formality. If Nest proposes a new teams table and you already have an organizations table that should be reused, say so before it builds. Fixing a plan takes seconds. Unwinding the wrong schema takes much longer.

Every change is a version

Each agent run is saved as a version, just like a normal edit. If you do not like the result, roll back with one click and your app returns to exactly where it was, database migrations included. Versions also sync to GitHub if you have two-way sync turned on, so your team can review an agent’s work as an ordinary commit.

Your code stays yours

Agent mode writes the same TypeScript you would get from any other Nest change. You can read it, edit it by hand or export it at any time. As always, your prompts and code are never used to train models.

Where it helps most

We built Agent mode for the changes that used to take a few rounds of back and forth:

• Adding a feature that touches the database, the API and several pages at once.

• Wiring up an integration such as Stripe, Slack or Okta, where a missed step only shows up at runtime.

• Refactoring a flow that already works, where the risk is breaking something you were not looking at.

For a quick copy tweak or a color change, the regular editor is still faster, and visual edits are faster still.

Try it

Open any project and switch the composer from Edit to Agent. Start with something you would normally ask for in three or four messages and see how far one request gets you. If you are new to Nest, our guide to writing a first prompt is a good place to begin, and the pricing page lists the agent allowance on each plan.

We will keep sharing what we learn as more people use it. If a run surprises you, good or bad, send it our way from the version menu. Every report is read by the team that built it.