Advanced Settings

The Advanced tab contains the agent state, category access, response limits, and external-channel behavior. Open Agents, select the agent, and choose Advanced in the sidebar.

Changes remain a draft. Use Save changes to apply them or Discard to return to the last loaded values. If saving reports an error, check both the fields and category before trying again: those two groups are persisted by separate operations.

Initial agent values

An agent created from the current Agents page starts with these values:

Setting	Initial value	Available range
Agent active	on	on or off
Temperature	1.0	0.0 to 2.0, in 0.1 steps
Context memory	30 messages	1 to 50
Search earlier conversations	on	on or off
Maximum reply length	4096 tokens	1024 to 16384, in 1024 steps
Maximum tool calls	14	1 to 20
Group messages	off	off or a 10- to 60-second window
Split response into multiple messages	off	on or off

Caution

The current interface help text still uses 0.7, 10, and 50 as reference values for some of these options. The table above describes the agent actually created by the page and the ranges enforced by the editor and runtime today.

State and category

Agent active controls whether the agent can reply. After you turn it off and save, the agent stops replying through channels, automations, and triggers, but keeps its configuration. Turn it back on and save to use it again.

Category limits access to the agent according to the categories authorized for each team. No category removes this category filter; organization permissions still apply. If the expected category is unavailable, ask an authorized person to review categories and teams instead of selecting a different one only to complete the save.

Response behavior

Temperature

Controls output variety. Values near 0.0 favor more predictable replies; high values increase variation. Start low for support, extraction, and procedures. Increase gradually for ideation and creative writing, and validate with real replies.

Context memory

Defines how many recent messages from the conversation enter the next turn’s context. A larger window helps with long conversations but sends more content to the model. Reduce it only when older context is unnecessary.

Search earlier conversations

When enabled, this gives the agent a tool that can locate summaries of the same person’s earlier conversations with the same agent. It works only in an individual Hub conversation: it does not cross people, agents, rooms, person-to-person direct chats, or external channels. Search results contain title, date, topics, and a summary, not the full transcript.

This option differs from Context memory: memory reads recent messages in the current conversation, while search looks through summaries of other eligible conversations.

Maximum reply length

Defines the configured reply budget. Lower values tend to reduce length, cost, and latency; higher values allow longer answers. Reasoning models may receive an additional technical reserve to think before answering, so do not treat this control as an exact word count.

Maximum tool calls

Limits how many tools the agent can chain in one turn before its final wrap-up. Each knowledge-base query, API call, transfer, or other tool counts as one call. The runtime reserves one additional step for the final answer and clamps legacy values above 20 to 20. The turn’s total time limit can still stop a flow before it reaches the selected number.

Replies on external channels

Group messages

When you enable Group messages, choose a window from 10 to 60 seconds. WhatsApp QR, Z-API, Uazapi, Evolution, Telegram, and the async API wait for the sequence and process incoming messages as one request. Each new message resets the wait, with a ceiling of three times the window measured from the first message.

When disabled, those channels still pass through the serialization queue, but each message forms its own group and is processed without the wait window. This setting does not apply to the Hub, public page, widget, Instagram Direct, WhatsApp Official, or the sync API.

Split response into multiple messages

When enabled, splitting turns each paragraph separated by a blank line into an independent delivery. It applies to Telegram, Instagram Direct, WhatsApp Official, and WhatsApp QR, Z-API, Uazapi, and Evolution connections. The Hub, public page, widget, and API ignore the setting.

Example of a reply delivered as three messages:

Sure, I can help.

2 PM is available.

Would you like me to confirm it?

Markdown images in the form ![description](https://example.com/image.jpg) are extracted and delivered as images. Audio, video, and files do not receive this treatment. Between parts, the channel may show a typing indicator and applies a short pause; do not promise an exact cadence to the customer.

Delete Agent

The Danger Zone appears only to people with the agents.delete permission. The visible button is Delete Agent, but the operation retains the record and message history for analytics: it marks the agent as archived and inactive, deactivates active triggers, and removes its team links.

The dialog shows how many triggers will be deactivated and how many teams will lose the agent. To confirm, copy or type the agent name exactly, including capitalization and spaces, then choose Delete permanently.

The current interface has no restore action. Before confirming, record any configuration that must be rebuilt and check the effect on journeys that depend on the agent.

Validation checklist

After saving:

1. leave the tab and return to confirm the persisted values;
2. test one short reply and one that uses tools;
3. if you changed memory or earlier-conversation search, validate an individual Hub conversation;
4. if you changed grouping or splitting, test the matching external channel;
5. check cost, timing, and format before releasing the agent to customers.

Continue with System Prompts, Models, and Testing Your Agent to validate the complete configuration.