Reference
Agents
On this page
Agents are LLM-backed participants in a channel’s activities. They read the ticket context and conversation history, then respond with a comment.
Where agents come from
An agent belongs to your organization. Create it under Conclude Agents > Agents in the Admin console. It posts under its own identity. Deploy it to a channel to let it respond in that channel and on the channel’s app tickets. The same agent can be deployed to several channels.
The Deployment page under Conclude Agents lists channels with deployed agents, including agents set to Off. Select a channel to add or remove agents, set their modes and configure learning. Select Save to apply your changes. Deploy agents lets you add several agents to one channel at a time.
The app editor’s Agents tab has the same controls for the app’s channel. Saving there saves the app first, then the agent settings. An agent’s own Deployment tab shows its deployments in a read-only list and links to the Deployment page.
Conclude’s agents
Conclude supports two types of agents for background tasks:
- A learning agent summarizes channel tickets and conversations.
- An agent creator generates new agents from your descriptions.
Create these with Create agent and edit them like other agents. Conclude generates their instructions from your description and the content they will process. For a learning agent, this includes the app’s YAML without its lookup configuration and scripts. Use the default instructions starts from a template instead.
These agents do not reply in channels or use knowledge collections. Their instructions define the task, and Conclude adds the required output format to the prompt. Each uses its configured model and API key, and its model calls count toward agent usage.
Without a configured agent for the task, Conclude uses its built-in instructions. Learning uses the organization’s key for the model provider; agent generation uses the included quota. An organization can have one agent creator and any number of learning agents. Each learning agent receives a randomly selected name.
Agents on tickets
Agents deployed to an app’s channel receive new tickets, comments, file uploads and incoming emails. Each agent responds according to its mode in that channel. They do not receive ticket edits, conclusions, button clicks, comments on closed tickets or comments from another app’s bot. Comments that the Conclude bot posts on behalf of external users, such as synced customer comments, are supported.
For an app with mode: channel, each ticket has its own channel. Agents deployed
to the app’s channel still respond on these tickets. A linked agent invited to a
ticket’s channel also receives its comments. It uses its mode in the app’s channel
if deployed there, or its default mode otherwise.
Agents are not configured in the app’s YAML. Conclude ignores legacy agent:
entries in triggers, along with their mode: and onMention: settings. The next
save removes these settings and any trigger entries left without an action.
To address an agent in a channel or on a ticket, type @ followed by its name.
For a linked agent, you can also @mention its user account. Agents without a
linked account do not appear in the mention list, so you must type the full name.
The mode
An agent’s mode in a channel applies to both channel conversations and the app’s tickets. Set the default mode on the agent’s General tab. To override it for a channel, select the channel on the Deployment page.
| Mode | When the agent responds |
|---|---|
| Mention | When a message, comment or new ticket @mentions it |
| Conversational | Also when addressed by name |
| Autonomous | It reviews every message, comment and new ticket and decides whether to respond |
| Off | It does not respond, but remains deployed |
For a new ticket, Conclude checks the title and description for mentions or text addressed to the agent. Conversational mode uses the same check as for comments.
Creating a ticket from a channel message, with 🎫 or the message action, does not notify agents again. They already received the original message, but they receive subsequent ticket comments.
Use Off to pause an agent in a channel its linked account cannot leave. For example, Microsoft Teams includes team members in every standard channel, so you cannot remove a linked account from just one of those channels.
An agent set to Off remains deployed. Conclude continues to store recent messages and receive channel events. When you turn the agent back on, it can read the stored conversation. Off also pauses replies on the channel’s app tickets.
What the agent sees
For each invocation, the agent receives:
- The app’s
context, or itsdescriptionwhen nocontextis set, as a system-prompt prefix. - The agent’s own instructions, from the admin console.
- Current ticket state — title, body, status, severity, owner, members, deadline, conclusion, and every custom attribute that has a value.
- The complete action history of the activity as alternating user /
assistant /
[system]messages (comments, owner changes, attribute changes, etc.). - Optional lookup data — if the app has a
lookup:endpoint configured for this event, the response body is included as additional context. - The triggering event description (e.g. “new comment from Alice”).
Learning
With learning enabled, Conclude stores summaries in a knowledge collection named “Learning: #channel”. It uses three sources:
| Source | Summary | When it is generated |
|---|---|---|
| A closed ticket from the channel’s app | A case file with the problem, cause and resolution | When the ticket closes |
| A thread in the channel | A case file with the question, answer and details | Two hours after the last reply; later replies cause an update |
| The day’s messages outside threads | A digest of topics, decisions and open questions | Just after midnight UTC |
The channel’s learning agent generates these summaries using its instructions, model and API key. Conclude defines the output format, while the agent’s instructions define the sections in ticket case files.
Deploying a learning agent enables message storage and adds the Conclude bot to the channel, even if no other agent is deployed there. Message text is kept for seven days instead of one hour. Turning learning off removes the stored messages and the learning agent’s deployment. Removing the app also turns learning off.
Connected channels learn only from tickets. A learning agent does not retain channel conversations there. Messages from the other organization are removed within minutes, as in a channel without an agent.
Conclude skips threads with only one participant, days with fewer than three messages and content the model identifies as having no useful information to retain.
If learning is enabled without an active learning agent, Conclude uses its built-in instructions and shows a warning in place of the agent. This can happen in channels that enabled learning before learning agents were introduced, after a learning agent is deleted or if Conclude could not install the agent.
New summaries appear on the Learning page for review. Agents can search
approved summaries with search_learning when responding on the app’s tickets or
in the channel. They do not search these summaries on connected tickets or in
connected channels, where people from another organization may have access.
To enable learning, select Learn from this channel in the channel’s drawer on the Deployment page or on the app editor’s Agents tab. You need permission to manage agents and the channel’s app. Choose an existing learning agent or generate one for the app. If the organization has no learning agent, you can use the default. The menu beside the agent lets you change it. Clear the checkbox to turn learning off.
Learning is available only for channels with an app. There is no learning setting in the app’s YAML.
Loop protection and rate limits
- Self-loop prevention. An agent never replies to its own comments. Comments authored by an agent are flagged, and a ticket’s dispatch skips an agent on its own comment.
- Cross-agent loop prevention. Any comment a script or an agent posts carries a marker that blocks further agent/script dispatch. The chain stops after one hop.
- Per-activity rate limit. Agent comments are throttled to 5 per minute per activity; excess is dropped.
- Per-activity comment cap. The agent’s max comments (or 20 by default) — the agent stops responding once agents have posted this many comments on a single activity, to prevent runaway threads.
Lookup-backed context
If the app has a lookup: endpoint, the agent handler POSTs to it with
{event, activity: {title, email}} and includes the response as extra
context in the prompt. Configure which events trigger lookup via the
on: list:
lookup:
url: https://crm.example.com/conclude-lookup
secret: crm_api_token # references a Secret by name
on: [onCreate, onComment]
timeout: 5s
See Connectors.