Publishing Checklist

Use this list after configuring the agent and before sharing a link, installing the Widget, or opening a channel to real users. Test common cases, scope limits, and expected failures—a correct answer to the first example does not prove that the flow is ready.

Save before testing

The Test agent panel uses the saved version. If changes are pending, save them before starting a new test conversation.

Prompt and security

• The objective, audience, and expected outcome are explicit in Prompt.
• The agent knows which information it must request before answering or performing an action.
• The prompt defines scope boundaries and what to do when a request cannot be fulfilled.
• The instructions explain when to use each tool and which data must never be invented.
• Security is configured when the flow needs additional protection against jailbreaks and prompt injection.
• Out-of-scope questions, conflicting instructions, and attempts to obtain inappropriate data have been tested.

Model, multimodal, and advanced settings

• The model selected in Model supports the capabilities the agent uses, such as tools, long context, images, or files.
• The input and output cost shown for the model fits the expected volume.
• In Multimodal, images, files, and audio are enabled and configured only when the channel needs to receive them.
• In Advanced, Context memory, Maximum reply length, and Maximum tool calls have been adjusted to the conversation type.
• Group messages and Split response into multiple messages have been validated in the target channel because they do not apply to every channel in the same way.
• Agent active is on. An inactive agent does not reply in channels, automations, or the test panel.

Tools and knowledge

• Every required tool is active in Tools, and its connection or credential is valid.
• Custom HTTP tools have been tested in their editor; other tools have been triggered in test conversations with valid and invalid inputs.
• The agent requests required parameters before calling a tool and explains what happened when the call fails.
• The prompt uses the correct names of connected tools and does not mention a disabled or nonexistent tool.
• The correct knowledge bases and context blocks are selected for the agent in Knowledge.
• Required base items have the Processed status, with no relevant items in Pending, Processing, or Error.
• Questions with a known answer and questions with no answer in the base have been tested to check retrieval, refusal, and any AutoLearn record.

Inboxes and human support

• In Inboxes, the inbox used for launch appears as Connected and points to the correct agent—or to Human support when that is intentional.
• The real channel has been tested end to end. A test in the editor does not replace sending through a public page, Widget, Telegram, WhatsApp, Instagram, or API.
• The welcome message has been reviewed for the public page or Widget when that field is used.
• The inbox schedule defines who answers in each period, who answers outside business hours, and exceptions for special dates. The organization time zone has also been checked.
• The human-support path has been tested through the inbox schedule/destination or the Transfer Conversation tool.
• If the channel accepts attachments, expected file types and sizes have been tested against the agent’s Multimodal configuration.

See channel configuration in Inboxes.

Verification after testing

• The external conversation appears in Conversations with the expected inbox, contact, and agent.
• Human transfer, AI reactivation, and conversation closure work according to the team’s process.
• In Analytics → Agents, the owner knows how to open the agent and monitor volume, tool usage, and tool errors.
• In Analytics → Credits, consumption by agent and period will be monitored after launch.
• Executions will be used only to monitor Schedules and Follow-ups; it is not the general log for tool calls.
• Someone is responsible for reviewing conversations, correcting failures, and updating the prompt, tools, and knowledge after launch.

Once every applicable item has been validated, release the channel to a small group, monitor the first conversations, and only then expand the audience.