Try it: A Sandbox Chat with Your Agent

Try it: a sandbox chat with your agent

Try it is the sandbox surface for every agent in Macha. It lives as its own tab on every agent's detail page, and it lets you talk to the agent exactly the way a customer or teammate eventually will — but privately, on your own account, before the agent is wired into any real trigger or shared with anyone.

Think of it as the agent's rehearsal room. You're the only person in it. Every message you send exercises the real prompt, the real tools, and the real model. Nothing you type here reaches a customer.

Why we call it "Try it" and not "Chat"

The Try it tab is where you exercise an agent before you ship it. Naming it "Chat" made it look like a sidebar toy; naming it "Try it" nudges you to actually rehearse. If you're building an agent from scratch, this is where you spend most of your time between edits.

What Try it actually does

Every conversation you start in the Try it tab uses the same chat loop your customers will hit:

• Same system prompt as the agent will use in production, so what you see is what real users will see.
• Same tool inventory — every connector tool, custom API tool, sub-agent, and knowledge source assigned to the agent is available in the chat.
• Same model as the agent is configured to run on. Changing the model in the Configuration tab immediately affects new Try it conversations.
• Same confirmation gates — every write tool that requires confirmation still requires confirmation here. You approve or reject each one just like a customer would.

The difference from production is who is on the other side: you. No customer sees the conversation. Nothing lands in a shared queue. The agent doesn't need a trigger set up to be usable — Try it is available the moment an agent has a prompt and one tool.

Try it vs. Simulations vs. Agent Evaluation

Three related surfaces, easy to confuse:

• Try it is an interactive sandbox: you drive the conversation, one message at a time, and real tools execute (subject to confirmation gates on writes). Best for quick "does this even work?" checks and one-off rehearsals.
• Agent Simulations replay a batch of past records (Zendesk tickets today) through the agent with write tools stubbed. No writes ever fire. Best for previewing how the agent would behave across a realistic pool of tickets before turning the trigger on.
• Agent Evaluation grades conversations that already happened — real customers, real writes — after the fact. Best for auditing production quality.

Reach for Try it while you're building. Reach for Simulations before you flip the trigger. Reach for Evaluation once real traffic is flowing.

Making writes safe in Try it

Because Try it uses the real tools, a write tool called during a Try it conversation will actually write — unless the tool is marked as requiring confirmation, in which case a confirmation card appears in the chat and nothing happens until you approve it. Every connector tool that mutates external state (sending a reply on a ticket, updating a status, creating an order) ships with confirmation on by default.

Two tips for safer rehearsals:

• Reject the confirmation to see what the agent would have written without the write firing. The tool result comes back marked as rejected; the agent can react to that in the next turn.
• Point the agent at test records when you can. If the connector supports a sandbox account (e.g. a second Zendesk instance for staging), assign that instance in Configuration and Try it will use it. Multiple instances of the same connector are natively supported.

Sharing a Try it conversation with a teammate

Every Try it conversation gets a URL. Copy it from the address bar and share it with a teammate on the same organization — they'll see the full transcript, including tool calls, confirmations, and the model's reasoning if that model exposes it. It's the fastest way to say "look, this is what the agent did on this input" without a screen share.

Starting a Try it conversation

Open any agent, click the Try it tab, and start typing. The first message spawns a new conversation; from then on, use the New chat button in the sidebar to start a fresh one. The History tab shows every Try it conversation you and your teammates have run against this agent, so you can revisit old rehearsals to compare behaviour after a prompt change.

Credit cost

Try it conversations use the same credit accounting as production runs. Each assistant reply costs the credit price of the model the agent is on. Free credits on the trial cover a healthy amount of rehearsal — a Sonnet 5 conversation with a few tool calls typically lands under ten credits.

What Try it is not

• It's not a dry run. Writes with confirmation off will fire for real. If you're not sure whether a tool is safe, use Simulations first.
• It's not read-only preview mode. The agent behaves exactly the way it will in production, including calling out to LLM providers, running knowledge search, and hitting connector APIs.
• It's not versioned. If you change the prompt mid-conversation, the next message uses the new prompt. If you want to preserve a comparison, save a snapshot in Configuration first.