Skip to main content
Try it is a chat next to the Agent tab’s canvas. Each message you send runs your whole agent once, live, on the real model, and the reply is the answer your agent returns. It is for one input at a time, while you build. To run a set of test cases and grade every answer, set up evals instead. Owner only Every run spends your own model key, so people you share the session with see the canvas but not Try it.

​
Open Try it

Try it sits on the toolbar in the bottom corner of the canvas, beside the evals button. Its tooltip reads “Run the whole flow on your own input”. The button shows whenever the canvas draws your agent, which happens once the session’s agent has built one. The chat opens beside the canvas, or below it on a narrow workspace. Drag the divider to give either more room, and double-click it to reset. To close the chat, click Try it again or Close the chat in its header.

​
Give it an input and run

The composer has one field for each input your agent declares, named in words, with its type beside the name and its description as the placeholder.
1

Fill the fields

Click a card under Start from a sample to fill every field without sending. If there are no samples, the chat says “No sample inputs for this agent yet. Write your own below, or have the builder draft a few.” Click Ask the agent for samples and the session’s agent writes three to five. After your first message, samples move to a row of pills above the composer, shown while the fields are empty or hold a sample. A check marks the ones you’ve already sent in this chat.
2

Pick the framework, if there is a choice

The framework chip shows what this message runs on. If your agent was built for more than one framework, click it to choose. A framework whose live check hasn’t passed reads ”· not proved live” and still runs. The model chip beside it names the model, and its tooltip names the key the run reads, for example “The run reads OPENAI_API_KEY”.
3

Run it

Click Run, or press Enter in a text field. Shift + Enter adds a new line. You need to fill at least one field, unless the agent takes no inputs. In that case the composer says “This agent declares no inputs. Run sends it as it is.”
Fields you leave empty take the default values from the agent’s inputs.json. If an input has no value in either place, the run is refused with a line naming the inputs it needs. Each message is a fresh run: your agent sees nothing from earlier messages in the chat. To keep the context of a conversation, put it in the input. Only one message runs at a time. If you press Run while an earlier message is still running, the button reads Queued. Your message then runs by itself, with the fields as they are at that point. Click Queued to cancel it. The queue lives in the page, so closing the tab drops it.

​
Watch the run

While a message runs, the canvas shows where it is. Each card gets a dot: pulsing while that step runs, green when it is done and red if it failed. The Inputs card turns green as the run starts, and the Answer card gets its dot when the run ends. The strip under the chat header shows the same dots in run order. Steps that run together are joined by +. The run takes place in the session’s machine and doesn’t depend on your tab. If you close the chat or the tab, it keeps going. Open the chat again and Alpha picks the run back up. Agents built on the Claude Agent SDK don’t report their steps, so on those only the Inputs and Answer cards light up.

​
Read the answer

When the run ends, the card under your message shows what your agent returned. The header shows the status (Done, Did not finish or Stopped) and how many seconds the run took. Short fields, such as numbers, flags or a word or two, appear as tiles. Longer text follows as sections under each field’s name. Two other things can appear on an answer card:
  • If the run passed but returned nothing, the card says “The run finished without an answer.”
  • An earlier version tag means the agent has been checked again since that message ran, so the message ran an older build.
Use New chat to start a clean chat and Earlier chats to reopen an old one.

​
What if a run fails or won’t start?

When a run fails, the card says “This run did not finish.” Click Details to see the run’s own reason, which reads “No reason was recorded.” if there was none. Click Ask the builder to look into it to send the failure to the session’s agent. It reads the run’s logs, explains the cause and shows you a fix before it edits anything. “This run ended early: the sandbox stopped under it. Send your message again to try once more.” means the session’s machine stopped while the run was going. Nothing is wrong with your agent, so send the message again. If something is missing before you start, the chat says what it is, usually with one button that fixes it.

​
What does a run cost, and what is kept?

Every message is a real run on the real model, and it is billed to the key in your secrets, not to Alpha. An agent with several steps can make several model calls per message.
The session-wide limits are on Limits.
Every chat and run is saved in your agent’s folder, under eval-runs/.playground/, and you can open it in Files. When a run ends, Alpha replaces any environment value of 8 characters or more whose variable name ends in KEY, TOKEN, SECRET or PASSWORD with [redacted] in those files. Your secrets are also hidden in everything the chat shows. While a run is going, only step names, statuses and timings leave the machine.

​
Next steps

Set up evals

Turn the inputs you tried by hand into graded test cases that run in one go.

The Agent tab

Read each step’s instructions, model and tools in the inspector before you run it.

Secrets

Add the model key your agent reads, for this session or for every session.

Build an AI agent

Change what the agent takes or returns, then come back and try it again.