Chat orchestration

One conversation runs the environment

Chat orchestration is the conversation on the left of every environment’s workspace. It is how you keep working with the builder, the Future AGI agent that read your agent and built the environment: you ask in plain language, and it changes the environment for you.

You don’t open a scenario editor to add ten scenarios, or hunt for the field that holds a tool’s arguments. You say what you want in the Reply to the builder… box, under the heading Ask, correct, or steer what it builds, and the builder does the work and reports back. Everything you ask for is kept in the conversation, so the history of how the environment came to be sits next to the environment itself. Collapse chat folds the conversation into a slim rail when you want the tabs at full width, and Expand chat brings it back.

What you see while it works

Each step the builder takes shows up as a line under its reply:

• a green dot means the step succeeded
• a red dot means one of the builder’s own checks turned the step away; it needs nothing from you, because the builder fixes the problem and tries again

The reply itself reports only what was kept. A scenario written through the chat is checked before it is kept, the same as one written at build time, so a red dot is a problem caught before it reached your environment.

When the builder genuinely needs something only you know, such as which of two meanings you intended, it asks, and the question appears in the conversation for you to answer. While it is working, the box reads Working on your last message…, and you can stop the turn.

What it can change

Chat orchestration changes the test environment. For example:

Ask for	Such as
An answer about the build, a run or a scenario	”What stage is the build at?” or “Why does this scenario expect a transfer?”
More scenarios, past the count you set at build	”Add five scenarios where the caller is in a hurry”
A scenario removed, or replaced with a harder one	”Replace the plain refill scenario with one where the caller corrects their date of birth”
A different caller	”Make this caller impatient and give them a different account number”
A contract correction	”The refund tool needs the order ID, not the email address”

A contract correction changes the environment’s understanding of your agent, so the scenarios and grading that follow use the corrected version, and the contract keeps the correction apart from what was read from source. Review the contract covers corrections step by step. While a run is going, its progress also appears in the conversation under Background run activity, and you can ask about it there.

What it refuses

The line is simple: chat orchestration never changes the agent under test, or how the environment reaches it. It refuses, straight away and without starting work, to:

• change the phone number a simulation dials, or the assistant ID, agent ID, URL or credentials it connects with
• change the agent’s model or voice settings
• edit or deploy the agent’s prompt, instructions, source code or platform configuration

Make those changes where your agent lives, on its hosted platform or in its code, then run a simulation against it. The chat can still help on either side of that: it can draft a prompt change for you to apply yourself, and add a scenario that tests whether the change worked.

When it is available

The conversation opens once the environment shows Ready. While it is being built, the box is locked and reads Environment is still being built, and the workspace fills in as each stage of the build lands. Results from runs that already finished are never changed by a request; edits apply to the runs that come after. If a request is refused or the chat stays locked, Environment FAQ & fixes covers why.

Keep exploring